feat: manipulate ai_surrogate_species

CMakeLists.txt

@@ -16,7 +16,6 @@ list(APPEND CMAKE_MODULE_PATH "${POET_SOURCE_DIR}/CMake")
 get_poet_version()
 
 find_package(MPI REQUIRED)
-
 find_package(RRuntime REQUIRED)
 
 add_subdirectory(src)

README.md

@@ -98,6 +98,11 @@ following available options:
   slowed down significantly. Defaults to _OFF_.
 - **POET_PREPROCESS_BENCHS**=*boolean* - enables the preprocessing of
   predefined models/benchmarks. Defaults to *ON*.
+- **USE_AI_SURROGATE**=*boolean* - includes the functions of the AI
+  surrogate model. When active, CMake relies on `find_package()` to find
+  an a implementation of `Threads` and a Python environment where Numpy
+  and Keras need to be installed. Defaults to _OFF_.
+
 
 ### Example: Build from scratch
 
@@ -231,37 +236,78 @@ mpirun -n 4 ./poet --dht --dht-snaps=2 barite_het_rt.R barite_het.rds output
 
 ### Example: Preparing Environment and Running with AI surrogate
 
-To run the AI surrogate, you need to install the R package `keras3`. The
-compilation process of POET remains the same as shown above.
+To run the AI surrogate, you need to have a Keras installed in your
+Python environment. The implementation in POET is agnostic to the exact
+Keras version, but the provided model file must match your Keras version.
+Using Keras 3 with `.keras` model files is recommended. The compilation
+process of POET remains mostly the same as shown above, but the CMake
+option `-DUSE_AI_SURROGATE=ON` must be set.
 
-In the following code block, the installation process on the Turing Cluster is
-shown. `miniconda` is used to create a virtual environment to install
-tensorflow/keras. Please adapt the installation process to your needs.
+To use the AI surrogate, you must declare several values in the R input
+script. This can be either done directly in the input script or in an
+additional file. This file can be provided by adding the file path as the
+element `ai_surrogate_input_script` to the `chemistry_setup` list in the
+R input script.
 
-<!-- Start an R interactive session and install the required packages: -->
+The following variables and functions must be declared:
+- `model_file_path` [*string*]: Path to the Keras model file with which 
+the AI surrogate model is initialized. 
 
-```sh
-# First, install the required R packages
-R -e "install.packages('keras3', repos='https://cloud.r-project.org/')"
+- `validate_predictions(predictors, prediction)` [*function*]: Returns a
+boolean vector of length `nrow(predictions)`. The output of this function
+defines which predictions are considered valid and which are rejected. 
+the predictors and predictions are passed in their original original (not
+transformed) scale. Regular simulation will only be done for the rejected
+values. The input data of the rejected rows and the respective true results
+from simulation will be added to the training data buffer of the AI surrogate
+model. Can eg. be implemented as a mass balance threshold between the
+predictors and the prediction.
 
-# manually create a virtual environment to install keras/python using conda, 
-# as this is somehow broken on the Turing Cluster when using the `keras::install_keras()` function
-cd poet
 
-# create a virtual environment in the .ai directory with python 3.11
-conda create -p ./.ai python=3.11
-conda activate ./.ai
+The following variables and functions can be declared:
+- `batch_size` [*int*]: Batch size for the inference and training functions,
+defaults to 2560.
 
-# install tensorflow and keras
-pip install keras tensorflow[and-cuda]
+- `training_epochs` [*int*]: Number of training epochs with each training data
+ set, defaults to 20.
  
-# add conda's python path to the R environment
-# make sure to have the conda environment activated
-echo -e "RETICULATE_PYTHON=$(which python)\n" >> ~/.Renviron
-```
+- `training_data_size` [*int*]: Size of the training data buffer. After
+the buffer has been filled, the model starts training and removes this amount
+of data from the front of the buffer. Defaults to the size of the Field.
+
+- `use_Keras_predictions` [*bool*]: Decides if the Keras prediction function
+should be used instead of the custom C++ implementation (Keras might be faster
+for larger models, especially on GPU). Defaults to false.
+
+- `use_k_means_clustering` [*bool*]: Decides if the K-Means clustering function
+will be used to separate the field in a reactive and a non-reactive cluster.
+Training and inference will be done with separate models for each cluster.
+Defaults to false.
+
+- `model_reactive_file_path` [*string*]: Path to the Keras model file with
+ which the AI surrogate model for the reactive cluster is initialized. If 
+ ommitted, the models for both clusters will be initialized from
+ `model_file_path`
+
+- `disable_training` [*bool*]: Deactivates the training functions. Defaults to
+false.
+
+- `train_only_invalid` [*bool*]: Use only the data from PHREEQC for training
+instead of the whole field (which might contain the models own predictions).
+Defaults to false. 
+
+- `save_model_path` [*string*]: After each training step the current model
+is saved to this path as a .keras file.
+
+- `preprocess(df)` [*function*]: 
+Returns the scaled/transformed data frame. The default implementation uses no
+scaling or transformations.
+
+- `postprocess(df)` [*function*]: 
+Returns the rescaled/backtransformed data frame. The combination of preprocess()
+and postprocess() is expected to be idempotent. The default implementation uses
+no scaling or transformations.
 
-After setup the R environment, recompile POET and you're ready to run the AI
-surrogate.
 
 ```sh
 cd <installation_dir>/bin
@@ -312,44 +358,3 @@ important information from the OpenMPI Man Page:
 For example, on platforms that support it, the clock_gettime()
 function will be used to obtain a monotonic clock value with whatever
 precision is supported on that platform (e.g., nanoseconds).
-
-## Additional functions for the AI surrogate
-
-The AI surrogate can be activated for any benchmark and is by default
-initiated as a sequential keras model with three hidden layer of depth
-48, 96, 24 with relu activation and adam optimizer. All functions in
-`ai_surrogate_model.R` can be overridden by adding custom definitions
-via an R file in the input script. This is done by adding the path to
-this file in the input script. Simply add the path as an element
-called `ai_surrogate_input_script` to the `chemistry_setup` list.
-Please use the global variable `ai_surrogate_base_path` as a base path
-when relative filepaths are used in custom funtions.
-
-**There is currently no default implementation to determine the
-validity of predicted values.** This means, that every input script
-must include an R source file with a custom function
-`validate_predictions(predictors, prediction)`. Examples for custom
-functions can be found for the barite_200 benchmark
-
-The functions can be defined as follows:
-
-`validate_predictions(predictors, prediction)`: Returns a boolean
-index vector that signals for each row in the predictions if the
-values are considered valid. Can eg. be implemented as a mass balance
-threshold between the predictors and the prediction.
-
-`initiate_model()`: Returns a keras model. Can be used to load
-pretrained models.
-
-`preprocess(df, backtransform = FALSE, outputs = FALSE)`: Returns the
-scaled/transformed/backtransformed dataframe. The `backtransform` flag
-signals if the current processing step is applied to data that's
-assumed to be scaled and expects backtransformed values. The `outputs`
-flag signals if the current processing step is applied to the output
-or tatget of the model. This can be used to eg. skip these processing
-steps and only scale the model input.
-
-`training_step (model, predictor, target, validity)`: Trains the model
-after each iteration. `validity` is the bool index vector given by
-`validate_predictions` and can eg. be used to only train on values
-that have not been valid predictions.

R_lib/ai_surrogate_model.R

@@ -1,75 +0,0 @@
-## This file contains default function implementations for the ai surrogate.
-## To load pretrained models, use pre-/postprocessing or change hyperparameters
-## it is recommended to override these functions with custom implementations via
-## the input script. The path to the R-file containing the functions mus be set
-## in the variable "ai_surrogate_input_script". See the barite_200.R file as an
-## example and the general README for more information.
-
-require(keras3)
-require(tensorflow)
-
-initiate_model <- function() {
-  hidden_layers <- c(48, 96, 24)
-  activation <- "relu"
-  loss <- "mean_squared_error"
-
-  input_length <- length(ai_surrogate_species)
-  output_length <- length(ai_surrogate_species)
-  ## Creates a new sequential model from scratch
-  model <- keras_model_sequential()
-
-  ## Input layer defined by input data shape
-  model %>% layer_dense(units = input_length,
-                        activation = activation,
-                        input_shape = input_length,
-                        dtype = "float32")
-
-  for (layer_size in hidden_layers) {
-    model %>% layer_dense(units = layer_size,
-                          activation = activation,
-                          dtype = "float32")
-  }
-
-  ## Output data defined by output data shape
-  model %>% layer_dense(units = output_length,
-                        activation = activation,
-                        dtype = "float32")
-
-  model %>% compile(loss = loss,
-                    optimizer = "adam")
-  return(model)
-}
-
-gpu_info <- function() {
-  msgm(tf_gpu_configured())
-}
-
-prediction_step <- function(model, predictors) {
-  prediction <- predict(model, as.matrix(predictors))
-  colnames(prediction) <- colnames(predictors)
-  return(as.data.frame(prediction))
-}
-
-preprocess <- function(df, backtransform = FALSE, outputs = FALSE) {
-  return(df)
-}
-
-postprocess <- function(df, backtransform = TRUE, outputs = TRUE) {
-  return(df)
-}
-
-set_valid_predictions <- function(temp_field, prediction, validity) {
-  temp_field[validity == 1, ] <- prediction[validity == 1, ]
-  return(temp_field)
-}
-
-training_step <- function(model, predictor, target, validity) {
-  msgm("Training:")
-
-  x <- as.matrix(predictor)
-  y <- as.matrix(target[colnames(x)])
-
-  model %>% fit(x, y)
-
-  model %>% save_model_tf(paste0(out_dir, "/current_model.keras"))
-}

R_lib/ai_surrogate_model_functions.R

@@ -0,0 +1,30 @@
+## This file contains default function implementations for the ai surrogate.
+## To use pre-/postprocessing it is recommended to override these functions
+## with custom implementations via the input script. The path to the R-file
+## See the barite_50.R file as an example and the general README for more
+## information.
+
+preprocess <- function(df) {
+  return(df)
+}
+
+postprocess <- function(df) {
+  return(df)
+}
+
+set_valid_predictions <- function(temp_field, prediction, validity) {
+  temp_field[validity == 1, ] <- prediction[validity == 1, ]
+  return(temp_field)
+}
+
+get_invalid_values <- function(df, validity) {
+  return(df[validity == 0, ])
+}
+
+set_field <- function(temp_field, columns, rows, column_name_limit,
+                      byrow = FALSE) {
+  temp_field <- matrix(temp_field, nrow = rows, byrow = byrow)
+  temp_field <- setNames(data.frame(temp_field), columns)
+  temp_field <- temp_field[column_name_limit]
+  return(temp_field)
+}

R_lib/kin_r_library.R

@@ -76,9 +76,10 @@ master_iteration_end <- function(setup, state_T, state_C) {
             state_C <- data.frame(state_C, check.names = FALSE)
 
             ai_surrogate_info <- list(
-                prediction_time = if (exists("ai_prediction_time")) as.integer(ai_prediction_time) else NULL,
-                training_time = if (exists("ai_training_time")) as.integer(ai_training_time) else NULL,
-                valid_predictions = if (exists("validity_vector")) validity_vector else NULL
+                prediction_time = if (exists("ai_prediction_time")) ai_prediction_time else NULL,
+                predictions_validity = if (exists("validity_vector")) validity_vector else NULL,
+                cluster_labels = if (exists("cluster_labels")) cluster_labels else NULL,
+                predictions = if (exists("predictions")) predictions else NULL
             )
 
             SaveRObj(x = list(

bench/barite/barite_200.R

@@ -47,8 +47,7 @@ dht_species <- c(
 )
 
 chemistry_setup <- list(
-  dht_species = dht_species,
-  ai_surrogate_input_script = "./barite_200ai_surrogate_input_script.R"
+  dht_species = dht_species
 )
 
 # Define a setup list for simulation configuration

bench/barite/barite_200ai_surrogate_input_script.R

@@ -1,48 +0,0 @@
-## load a pretrained model from tensorflow file
-## Use the global variable "ai_surrogate_base_path" when using file paths
-## relative to the input script
-initiate_model <- function() {
-  init_model <- normalizePath(paste0(ai_surrogate_base_path,
-                                     "model_min_max_float64.keras"))
-  return(load_model_tf(init_model))
-}
-
-scale_min_max <- function(x, min, max, backtransform) {
-  if (backtransform) {
-    return((x * (max - min)) + min)
-  } else {
-    return((x - min) / (max - min))
-  }
-}
-
-preprocess <- function(df, backtransform = FALSE, outputs = FALSE) {
-  minmax_file <- normalizePath(paste0(ai_surrogate_base_path,
-                                      "min_max_bounds.rds"))
-  global_minmax <- readRDS(minmax_file)
-  for (column in colnames(df)) {
-    df[column] <- lapply(df[column],
-                         scale_min_max,
-                         global_minmax$min[column],
-                         global_minmax$max[column],
-                         backtransform)
-  }
-  return(df)
-}
-
-mass_balance <- function(predictors, prediction) {
-  dBa <- abs(prediction$Ba + prediction$Barite -
-             predictors$Ba - predictors$Barite)
-  dSr <- abs(prediction$Sr + prediction$Celestite -
-             predictors$Sr - predictors$Celestite)
-  return(dBa + dSr)
-}
-
-validate_predictions <- function(predictors, prediction) {
-  epsilon <- 3e-5
-  mb <- mass_balance(predictors, prediction)
-  msgm("Mass balance mean:", mean(mb))
-  msgm("Mass balance variance:", var(mb))
-  msgm("Rows where mass balance meets threshold", epsilon, ":",
-       sum(mb < epsilon))
-  return(mb < epsilon)
-}

bench/barite/barite_50ai_surr_mdl.R

@@ -1,18 +1,13 @@
 ## Time-stamp: "Last modified 2024-05-30 13:27:06 delucia"
 
-## load a pretrained model from tensorflow file
 ## Use the global variable "ai_surrogate_base_path" when using file paths
 ## relative to the input script
-initiate_model <- function() {
-    require(keras3)
-    require(tensorflow)
-    init_model <- normalizePath(paste0(ai_surrogate_base_path,
+model_file_path <- normalizePath(paste0(ai_surrogate_base_path,
                                         "barite_50ai_all.keras"))
-    Model <- keras3::load_model(init_model)
-    msgm("Loaded model:")
-    print(str(Model))
-    return(Model)
-}
+
+batch_size <- 1280
+training_epochs <- 20
+save_model_path <- "current_model.keras"
 
 scale_min_max <- function(x, min, max, backtransform) {
   if (backtransform) {
@@ -22,22 +17,105 @@ scale_min_max <- function(x, min, max, backtransform) {
   }
 }
 
-minmax <- list(min = c(H = 111.012433592824, O = 55.5062185549492, Charge = -3.1028354471876e-08, 
-                       Ba = 1.87312878574393e-141, Cl = 0, `S(6)` = 4.24227510643685e-07, 
-                       Sr = 0.00049382996130541, Barite = 0.000999542409828586, Celestite = 0.244801877115968),
-               max = c(H = 111.012433679682, O = 55.5087003521685, Charge = 5.27666636082035e-07, 
-                       Ba = 0.0908849779513762, Cl = 0.195697626449355, `S(6)` = 0.000620774752665846, 
-                       Sr = 0.0558680070692722, Barite = 0.756779139057097, Celestite = 1.00075422160624
-                       ))
+## Apply decimal logarithms handling negative values
+Safelog <- function (vec) {
+  rv <- range(vec)
+  if (max(abs(rv)) < 1) {
+    ret <- vec
+    ret[vec == 0] <- 0
+    ret[vec > 0] <- log10(vec[vec > 0])
+    ret[vec < 0] <- -log10(-vec[vec < 0])
+  } else {
+    ret <- log10(abs(vec))
+  }
+  ret
+}
+Safeexp <- function (vec) {
+  ret <- vec
+  ret[vec == 0] <- 0
+  ret[vec > 0] <- -10^(-vec[vec > 0])
+  ret[vec < 0] <- 10^(vec[vec < 0])
+  ret
+}
+
+##' @title Apply transformations to cols of a data.frame
+##' @param df named data.frame
+##' @param tlist list of function names
+##' @return data.frame with the columns specified in tlist and the
+##'     transformed numerical values
+##' @author MDL
+TransfList <- function (df, tlist) {
+  vars <- intersect(colnames(df), names(tlist))
+  lens <- sapply(tlist[vars], length)
+  n <- max(lens)
+  filledlist <- lapply(tlist[vars],
+                       function(x) 
+                         if (length(x) < n)
+                           return(c(x, rep("I", n - length(x))))
+                         else
+                           return(x))
+  tmp <- df[, vars]
+  for (i in seq_len(n)) {
+    tmp <- as.data.frame(sapply(vars, function(x)
+                                  do.call(filledlist[[x]][i], list(tmp[, x]))))
+  }
+  return(tmp)
+}
+
+##' This function applies some specified string substitution such as
+##' 's/log/exp/' so that from a list of "forward" transformation
+##' functions it computes a "backward" one
+##' @title Apply back transformations to cols of a data.frame
+##' @param df named data.frame
+##' @param tlist list of function names
+##' @return data.frame with the columns specified in tlist and the
+##'     backtransformed numerical values
+##' @author MDL
+BackTransfList <- function (df, tlist) {
+  vars <- intersect(colnames(df), names(tlist))
+  lens <- sapply(tlist[vars], length)
+  n <- max(lens)
+  filledlist <- lapply(tlist[vars],
+                       function(x)
+                         if (length(x) < n)
+                           return(c(x, rep("I", n - length(x))))
+                         else
+                           return(x))
+  rlist <- lapply(filledlist, rev)
+  rlist <- lapply(rlist, sub, pattern = "log", replacement = "exp")
+  rlist <- lapply(rlist, sub, pattern = "1p", replacement = "m1")
+  rlist <- lapply(rlist, sub, pattern = "Mul", replacement = "Div")
+  tmp <- df[, vars]
+  for (i in seq_len(n)) {
+    tmp <- as.data.frame(sapply(vars, function(x)
+                                  do.call(rlist[[x]][i], list(tmp[, x]))))
+  }
+  return(tmp)
+}
+
+tlist <- list("H" = "log1p", "O" = "log1p", "Charge" = "Safelog",
+              "Ba" = "log1p", "Cl" = "log1p", "S(6)" = "log1p",
+              "Sr" = "log1p", "Barite" = "log1p", "Celestite" = "log1p")
+
+minmaxlog <- list(min = c(H = 4.71860987935512, O = 4.03435069501537,
+                          Charge = -14.5337693764617, Ba = 1.87312878574393e-141,
+                          Cl = 0, `S(6)` = 4.2422742065922e-07,
+                          Sr = 0.00049370806741832, Barite = 0.000999043199940672,
+                          Celestite = 0.218976382406766),
+                  max = c(H = 4.71860988013054, O = 4.03439461483133,
+                          Charge = 12.9230752782909, Ba = 0.086989273200771,
+                          Cl = 0.178729802869381, `S(6)` = 0.000620582151722819,
+                          Sr = 0.0543631841661421, Barite = 0.56348209786429,
+                          Celestite = 0.69352422027466))
 
 preprocess <- function(df) {
   if (!is.data.frame(df))
     df <- as.data.frame(df, check.names = FALSE)
-    
-    as.data.frame(lapply(colnames(df),
-                         function(x) scale_min_max(x=df[x],
-                                                   min=minmax$min[x],
-                                                   max=minmax$max[x],
+  tlog <- TransfList(df, tlist)
+  as.data.frame(lapply(colnames(tlog),
+                        function(x) scale_min_max(x = tlog[x], 
+                                                  min = minmaxlog$min[x],
+                                                  max = minmaxlog$max[x],
                                                   backtransform = FALSE)),
                 check.names = FALSE)
 }
@@ -45,25 +123,26 @@ preprocess <- function(df) {
 postprocess <- function(df) {
   if (!is.data.frame(df))
     df <- as.data.frame(df, check.names = FALSE)
-    
-    as.data.frame(lapply(colnames(df),
+  ret <- as.data.frame(lapply(colnames(df),
                               function(x) scale_min_max(x = df[x],
-                                                   min=minmax$min[x],
-                                                   max=minmax$max[x],
+                                                        min = minmaxlog$min[x],
+                                                        max = minmaxlog$max[x],
                                                         backtransform = TRUE)),
                        check.names = FALSE)
+
+  BackTransfList(ret, tlist)
 }
 
-mass_balance <- function(predictors, prediction) {
-    dBa <- abs(prediction$Ba + prediction$Barite -
-               predictors$Ba - predictors$Barite)
-    dSr <- abs(prediction$Sr + prediction$Celestite -
-               predictors$Sr - predictors$Celestite)
+mass_balance <- function(x, y) {
+  dBa <- abs(y$Ba + y$Barite -
+             (x$Ba + x$Barite))
+  dSr <- abs(y$Sr + y$Celestite -
+             (x$Sr + x$Celestite))
   return(dBa + dSr)
 }
 
 validate_predictions <- function(predictors, prediction) {
-    epsilon <- 1E-7
+    epsilon <- 1E-5
     mb <- mass_balance(predictors, prediction)
     msgm("Mass balance mean:", mean(mb))
     msgm("Mass balance variance:", var(mb))
@@ -72,19 +151,3 @@ validate_predictions <- function(predictors, prediction) {
          sum(ret))
     return(ret)
 }
-
-training_step <- function(model, predictor, target, validity) {
-  msgm("Starting incremental training:")
-
-  ## x <- as.matrix(predictor)
-  ## y <- as.matrix(target[colnames(x)])
-
-  history <- model %>% keras3::fit(x = data.matrix(predictor),
-                                   y = data.matrix(target),
-                                   epochs = 10, verbose=1)
-
-  keras3::save_model(model,
-                     filepath = paste0(out_dir, "/current_model.keras"),
-                     overwrite=TRUE)
-  return(model)
-}

src/CMakeLists.txt

@@ -18,7 +18,55 @@ target_sources(POETLib
   Chemistry/SurrogateModels/ProximityHashTable.cpp
 ) 
 
-target_include_directories(POETLib PUBLIC "${CMAKE_CURRENT_SOURCE_DIR}")
+target_include_directories(
+ POETLib PUBLIC
+ "${CMAKE_CURRENT_SOURCE_DIR}"
+ "${CMAKE_CURRENT_BINARY_DIR}"
+)
+
+option(USE_AI_SURROGATE "Compiles the AI surrogate functions with Python.h integration" OFF)
+if(USE_AI_SURROGATE)
+  add_definitions(-DUSE_AI_SURROGATE)
+  message("Building with AI surrogate functions")
+  message(" -- Needs Python (with Numpy & Keras) and Threads libraries")
+
+  find_package(CUDA)
+  IF (${CUDA_FOUND})
+    message(" -- Setting TensorFlow CUDA path to: ${CUDA_TOOLKIT_ROOT_DIR}")
+  ELSE (${CUDA_FOUND})
+  message(" -- No CUDA installation found for TensorFlow  ")
+  ENDIF (${CUDA_FOUND})
+
+  # make sure to use the python installation from the conda environment
+  if(DEFINED ENV{CONDA_PREFIX})
+    set(Python3_EXECUTABLE "$ENV{CONDA_PREFIX}/bin/python3")
+  endif()
+  find_package(Python3 COMPONENTS Interpreter Development NumPy REQUIRED)
+  
+  find_package(Threads REQUIRED)
+  
+  set_source_files_properties(
+    Chemistry/SurrogateModels/AI_functions.cpp
+    PROPERTIES COMPILE_FLAGS "-O3 -march=native -mtune=native"
+  )
+
+  target_sources(POETLib
+   PRIVATE
+   Chemistry/SurrogateModels/AI_functions.cpp
+  )
+  
+  target_include_directories(POETLib PUBLIC
+   "${Python3_INCLUDE_DIRS}"
+   "${Python3_NumPy_INCLUDE_DIRS}"
+  )
+  
+  target_link_libraries(POETLib
+   PUBLIC "${Python3_LIBRARIES}"
+   PUBLIC Threads::Threads pthread
+  )
+endif() # USE_AI_SURROGATE
+
+
 target_link_libraries(
  POETLib
  PUBLIC RRuntime
@@ -27,6 +75,10 @@ target_link_libraries(
  PUBLIC MPI::MPI_C
 )
 
+# Define Python API version
+target_compile_definitions(POETLib PRIVATE NPY_NO_DEPRECATED_API=NPY_1_7_API_VERSION)
+
+
 include(FetchContent)
 FetchContent_Declare(
   cli11
@@ -80,10 +132,11 @@ target_compile_definitions(POETLib PUBLIC STRICT_R_HEADERS OMPI_SKIP_MPICXX)
 
 file(READ "${PROJECT_SOURCE_DIR}/R_lib/kin_r_library.R" R_KIN_LIB )
 file(READ "${PROJECT_SOURCE_DIR}/R_lib/init_r_lib.R" R_INIT_LIB)
-file(READ "${PROJECT_SOURCE_DIR}/R_lib/ai_surrogate_model.R" R_AI_SURROGATE_LIB)
+file(READ "${PROJECT_SOURCE_DIR}/R_lib/ai_surrogate_model_functions.R" R_AI_SURROGATE_LIB)
 
 configure_file(poet.hpp.in poet.hpp @ONLY)
 
+
 add_executable(poet poet.cpp)
 target_link_libraries(poet PRIVATE POETLib MPI::MPI_C RRuntime CLI11::CLI11)
 target_include_directories(poet PRIVATE "${CMAKE_CURRENT_BINARY_DIR}")

src/Chemistry/SurrogateModels/AI_Python_functions/keras_AI_surrogate.py

@@ -0,0 +1,49 @@
+import tensorflow as tf
+import numpy as np
+from sklearn.cluster import KMeans
+import os
+
+os.environ["TF_XLA_FLAGS"] = "--tf_xla_cpu_global_jit"
+os.environ["XLA_FLAGS"] = "--xla_gpu_cuda_data_dir=" + cuda_dir
+
+def k_means(data, k=2, tol=1e-6):
+    kmeans = KMeans(n_clusters=k, tol=tol)
+    labels = kmeans.fit_predict(data)
+    return labels
+
+def initiate_model(model_file_path):
+    print("AI: Model loaded from: " + model_file_path, flush=True)
+    model = tf.keras.models.load_model(model_file_path)
+    return model
+
+def prediction_step(model, model_reactive, x, cluster_labels, batch_size):
+    # Catch input size mismatches
+    model_input_shape = model.input_shape[1:]
+    if model_input_shape != x.shape[1:]:
+        print(f"Input data size {x.shape[1:]} does not match model input size {model_input_shape}",
+              flush=True)        
+
+    # Predict separately if clustering is used
+    if cluster_labels:
+        cluster_labels = np.asarray(cluster_labels, dtype=bool)
+        # Combine results
+        prediction = np.zeros_like(x)
+        prediction[cluster_labels] = model_reactive.predict(x[cluster_labels], batch_size)
+        prediction[~cluster_labels] = model.predict(x[~cluster_labels], batch_size)
+    else:
+        prediction = model.predict(x, batch_size)
+    return np.array(prediction, dtype=np.float64)
+
+
+def get_weights(model):
+    weights = model.get_weights()
+    return weights
+
+def training_step(model, x, y, batch_size, epochs, 
+                  output_file_path):
+    history = model.fit(x, y, 
+                        epochs=epochs,
+                        batch_size=batch_size)
+    if output_file_path:
+        model.save(output_file_path)
+    return history

src/Chemistry/SurrogateModels/AI_functions.cpp

@@ -0,0 +1,880 @@
+#include <iostream>
+#include <string>
+#include <cstring>
+#include <vector>
+#include <Python.h>
+#include <numpy/arrayobject.h>
+#include <Eigen/Dense>
+#include <thread>
+#include <mutex>
+#include <condition_variable>
+#include "AI_functions.hpp"
+#include "poet.hpp"
+
+using namespace std;
+
+namespace poet {
+
+/**
+ * @brief Loads the Python interpreter and functions  
+ * @param functions_file_path Path to the Python file where the AI surrogate
+ * functions are defined
+ * @return 0 if function was succesful
+ */
+int Python_Keras_setup(std::string functions_file_path, std::string cuda_src_dir) {
+  // Initialize Python functions
+  Py_Initialize();
+  // Import numpy functions
+  _import_array();
+  PyRun_SimpleString(("cuda_dir = \"" + cuda_src_dir + "\"").c_str()) ;
+  FILE* fp = fopen(functions_file_path.c_str(), "r");
+  int py_functions_initialized = PyRun_SimpleFile(fp, functions_file_path.c_str());
+  fclose(fp);
+  if (py_functions_initialized != 0) {
+    PyErr_Print();
+    throw std::runtime_error(std::string("AI surrogate Python functions could not be loaded." ) + 
+                             "Are tensorflow and numpy installed?");
+  }
+  return py_functions_initialized;
+}
+
+/**
+ * @brief Loads the user-supplied Keras model  
+ * @param model_file_path Path to a .keras file that the user must supply as
+ * a variable "model_file_path" in the R input script
+ * @return 0 if function was succesful
+ */
+int Python_Keras_load_model(std::string model, std::string model_reactive, bool use_clustering) {
+  // Acquire the Python GIL
+  PyGILState_STATE gstate = PyGILState_Ensure();
+
+  // Initialize Keras default model
+  int py_model_loaded = PyRun_SimpleString(("model = initiate_model(\"" + 
+                                             model + "\")").c_str());
+  if (py_model_loaded != 0) {
+    PyErr_Print();
+    throw std::runtime_error("Keras model could not be loaded from: " + model);
+  }
+
+  if (use_clustering) {
+    // Initialize second Keras model that will be used for the "reaction" cluster
+    py_model_loaded = PyRun_SimpleString(("model_reactive = initiate_model(\"" + 
+                                          model_reactive + "\")").c_str());
+    if (py_model_loaded != 0) {
+      PyErr_Print();
+      throw std::runtime_error("Keras model could not be loaded from: " + model_reactive);
+    }
+  }
+  // Release the Python GIL
+  PyGILState_Release(gstate);
+  return py_model_loaded;
+}
+
+
+/**
+ * @brief Calculates the euclidian distance between two points in n dimensional space
+ * @param a Point a
+ * @param b Point b
+ * @return The distance
+ */
+double distance(const std::vector<double>& a, const std::vector<double>& b) {
+    double sum = 0.0;
+    for (size_t i = 0; i < a.size(); ++i) {
+        sum += (a[i] - b[i]) * (a[i] - b[i]);
+    }
+    return sqrt(sum);
+}
+
+/**
+ * @brief Assigns all elements of a 2D-Matrix to the nearest cluster center point
+ * @param field 2D-Matrix with the content of a Field object
+ * @param clusters The vector of clusters represented by their center points
+ * @return A vector that contains the assigned cluster for each of the rows in field
+ */
+std::vector<int> assign_clusters(const std::vector<vector<double>>& field, const std::vector<vector<double>>& clusters) {
+  // Initiate a vector that holds the cluster labels of each row
+  std::vector<int> labels(field[0].size());
+ 
+ for (size_t row = 0; row < labels.size(); row++) {
+    // Get the coordinates of the current row
+    std::vector<double> row_data(field.size()); 
+    for (size_t column = 0; column < row_data.size(); column++) {
+      row_data[column] = field[column][row];
+    }
+    // Iterate over the clusters and check which cluster center is the closest
+    double current_min_distance = numeric_limits<double>::max();
+    int current_closest_cluster;
+    for (size_t cluster = 0; cluster < clusters.size(); cluster++) {
+      double cluster_distance = distance(row_data, clusters[cluster]);
+      if (cluster_distance < current_min_distance) {
+        current_min_distance = cluster_distance;
+        current_closest_cluster = cluster;
+      }
+    }
+    labels[row] = current_closest_cluster;
+  }
+  return labels;
+}
+
+/**
+ * @brief Calculates new center points for each given cluster by averaging the coordinates
+ * of all points that are assigen to it
+ * @param field 2D-Matrix with the content of a Field object
+ * @param labels The vector that contains the assigned cluster for each of the rows in field
+ * @param k The number of clusters
+ * @return The new cluster center points
+ */
+std::vector<vector<double>> calculate_new_clusters(const std::vector<std::vector<double>>& field,
+                                                   const vector<int>& labels, int k) {
+  size_t columns = field.size();
+  size_t rows = field[0].size();
+  std::vector<std::vector<double>> clusters(k, std::vector<double>(columns, 0.0));
+  vector<int> count(k, 0);
+
+  // Sum the coordinates of all points that are assigned to each cluster 
+  for (size_t row = 0; row < rows; row++) {
+    int assigned_cluster = labels[row];
+    for (size_t column = 0; column < columns; column++) {
+      clusters[assigned_cluster][column] += field[column][row];
+    }
+    count[assigned_cluster]++;
+    }
+
+  // Take the average of the summed coordinates
+  for (size_t cluster = 0; cluster < k; cluster++) {
+    if (count[cluster] == 0) continue;
+    for (size_t column = 0; column < columns; column++) {
+      clusters[cluster][column] /= count[cluster];
+    }
+  }
+  return clusters;
+}
+
+/**
+ * @brief Performs KMeans clustering for the elements of a 2D-Matrix
+ * @param field 2D-Matrix with the content of a Field object
+ * @param k The number of different clusters
+ * @param iterations The number of cluster update steps
+ * @return A vector that contains the assigned cluster for each of the rows in field
+ */
+std::vector<int> K_Means(std::vector<std::vector<double>>& field, int k, int iterations) {
+  // Initialize cluster centers by selecting random points from the field 
+  srand(time(0));
+  std::vector<vector<double>> clusters;
+  for (size_t i = 0; i < k; ++i) {
+    std::vector<double> cluster_center(field.size());
+    int row = rand() % field.size();
+    for (size_t column = 0; column < cluster_center.size(); column++) {
+      cluster_center[column] = field[column][row];
+    }
+    clusters.push_back(cluster_center);
+  } 
+
+  std::vector<int> labels;
+  
+  for (size_t iter = 0; iter < iterations; ++iter) {
+    // Get the nearest cluster for each row
+    labels = assign_clusters(field, clusters);
+    // Update each cluster center as the average location of each point assigned to it
+    std::vector<vector<double>> new_clusters = calculate_new_clusters(field, labels, k);
+    clusters = new_clusters;
+  }
+
+
+  // Always define the reactive cluster as cluster 1
+  // Interprete the reactive cluster as the one on the origin of the field
+  // TODO: Is that always correct?  
+  int reactive_cluster = labels[0];
+  if (reactive_cluster == 0) {
+    for (size_t i; i < labels.size(); i++) {
+          labels[i] = 1 - labels[i];
+    }
+  }
+  return labels;
+}
+
+
+/**
+ * @brief Converts the std::vector 2D matrix representation of a POET Field object to a numpy array
+ * for use in the Python AI surrogate functions
+ * @param field 2D-Matrix with the content of a Field object
+ * @return Numpy representation of the input vector
+ */
+PyObject* vector_to_numpy_array(const std::vector<std::vector<double>>& field) {
+  npy_intp dims[2] = {static_cast<npy_intp>(field[0].size()), 
+                      static_cast<npy_intp>(field.size())};
+  
+  PyObject* np_array = PyArray_SimpleNew(2, dims, NPY_FLOAT64);
+  double* data = static_cast<double*>(PyArray_DATA((PyArrayObject*)np_array));
+  // write field data to numpy array
+  for (size_t i = 0; i < field.size(); ++i) {
+    for (size_t j = 0; j < field[i].size(); ++j) {
+      data[j * field.size() + i] = field[i][j];
+    }
+  }
+  return np_array;
+}
+
+/**
+ * @brief Converts a Pyton matrix object to a std::vector vector
+ * @param py_matrix Pyobject that must be a 2D matrix
+ * @result Vector that can be used similar to the return value of the Field object
+ * Field.AsVector() method.
+ */
+std::vector<double> numpy_array_to_vector(PyObject* py_array) {
+  std::vector<double> result;
+  if (!PyArray_Check(py_array)) {
+    std::cerr << "The model's output is not a numpy array." << std::endl;
+    return result;
+  }
+  // Cast generic PyObject to PyArrayObject
+  PyArrayObject* np_array = reinterpret_cast<PyArrayObject*>(py_array);
+  
+  // Get shape
+  int numDims = PyArray_NDIM(np_array);
+  npy_intp* shape = PyArray_SHAPE(np_array);
+  if (numDims != 2) {
+    std::cerr << "The model's predictions are not a 2D matrix." << std::endl;
+    return result;
+  }
+
+  // Copy data into std::vector format
+  double* data = static_cast<double*>(PyArray_DATA(np_array));
+  npy_intp size = PyArray_SIZE(np_array);
+  result.resize(size);
+  std::copy(data, data + size, result.begin());
+  return result;
+}
+
+/**
+ * @brief Uses the Python Keras functions to calculate predictions from a neural network.
+ * @param x 2D-Matrix with the content of a Field object
+ * @param batch_size size for mini-batches that are used in the Keras model.predict() method
+ * @return Predictions that the neural network made from the input values x. The predictions are
+ * represented as a vector similar to the representation from the Field.AsVector() method
+ */
+std::vector<double> Python_Keras_predict(std::vector<std::vector<double>>& x, int batch_size,
+                                         std::vector<int>& cluster_labels) {
+  // Acquire the Python GIL
+  PyGILState_STATE gstate = PyGILState_Ensure();
+  // Prepare data for Python
+  PyObject* py_df_x = vector_to_numpy_array(x);
+
+  // Prepare cluster label vector for Python
+  PyObject* py_cluster_list = PyList_New(cluster_labels.size());
+  for (size_t i = 0; i < cluster_labels.size(); i++) {
+    PyObject* py_int = PyLong_FromLong(cluster_labels[i]);
+    PyList_SET_ITEM(py_cluster_list, i, py_int);
+  }
+  
+  // Get the model and inference function from the global python interpreter
+  PyObject* py_main_module = PyImport_AddModule("__main__");
+  PyObject* py_global_dict = PyModule_GetDict(py_main_module);
+  PyObject* py_keras_model = PyDict_GetItemString(py_global_dict, "model");
+  PyObject* py_inference_function = PyDict_GetItemString(py_global_dict, "prediction_step");
+  
+  // Get secod model if clustering is used
+  PyObject* py_keras_model_reactive = Py_None;;
+  if (cluster_labels.size() > 0) {
+    py_keras_model_reactive = PyDict_GetItemString(py_global_dict, "model_reactive");
+  }
+
+  // Build the function arguments as four python objects and an integer
+  PyObject* args = Py_BuildValue("(OOOOi)",
+      py_keras_model, py_keras_model_reactive, py_df_x, py_cluster_list, batch_size);
+
+  // Call the Python inference function
+  PyObject* py_predictions = PyObject_CallObject(py_inference_function, args);
+
+  // Check py_rv and return as 2D vector
+  std::vector<double> predictions = numpy_array_to_vector(py_predictions); 
+  
+  // Clean up
+  PyErr_Print(); 
+  Py_XDECREF(py_df_x);
+  Py_XDECREF(py_cluster_list);
+  Py_XDECREF(args);
+  Py_XDECREF(py_predictions);
+  
+  // Release the Python GIL
+  PyGILState_Release(gstate);
+  return predictions;
+}
+
+/**
+ * @brief Uses Eigen for fast inference with the weights and biases of a neural network 
+ * @param input_batch Batch of input data that must fit the size of the neural networks input layer
+ * @param model Struct of aligned Eigen vectors that hold the neural networks weights and biases.
+ * Only supports simple fully connected feed forward networks.
+ * @return The batch of predictions made with the neural network weights and biases and the data
+ * in input_batch
+ */
+Eigen::MatrixXd eigen_inference_batched(const Eigen::Ref<const Eigen::MatrixXd>& input_batch, const EigenModel& model) {
+    Eigen::MatrixXd current_layer = input_batch;
+
+    // Process all hidden layers
+    for (size_t layer = 0; layer < model.weight_matrices.size() - 1; ++layer) {
+        current_layer = (model.weight_matrices[layer] * current_layer);
+        current_layer = current_layer.colwise() + model.biases[layer];
+        current_layer = current_layer.array().max(0.0);
+    }
+
+    // Process output layer (without ReLU)
+    size_t output_layer = model.weight_matrices.size() - 1;
+    return (model.weight_matrices[output_layer] * current_layer).colwise() + model.biases[output_layer];
+}
+
+/**
+ * @brief Uses the Eigen representation of the two different Keras model weights for fast inference
+ * @param model The model for the non reactive cluster of the field (label 0)
+ * @param model_reactive The model for the non reactive cluster of the field (label 1)
+ * @param x 2D-Matrix with the content of a Field object
+ * @param batch_size size for mini-batches that are used in the Keras model.predict() method
+ * @param Eigen_model_mutex Mutex that locks the model during inference and prevents updaties from
+ * the training thread
+ * @param cluster_labels K-Means cluster label dor each row in the field
+ * @return Predictions that the neural network made from the input values x. The predictions are
+ * represented as a vector similar to the representation from the Field.AsVector() method
+ */
+std::vector<double> Eigen_predict_clustered(const EigenModel& model, const EigenModel& model_reactive,
+                                            std::vector<std::vector<double>>& x, int batch_size,
+                                            std::mutex* Eigen_model_mutex, std::vector<int>& cluster_labels) {
+    const int num_samples = x[0].size();
+    const int num_features = x.size();
+    if (num_features != model.weight_matrices[0].cols() || 
+        num_features != model_reactive.weight_matrices[0].cols()) {
+        throw std::runtime_error("Input data size " + std::to_string(num_features) + 
+            " does not match model input layer sizes" + std::to_string(model.weight_matrices[0].cols()) + 
+            " / " + std::to_string(model_reactive.weight_matrices[0].cols()));
+    }
+
+    // Convert input data to Eigen matrix
+    Eigen::MatrixXd full_input_matrix(num_features, num_samples);
+    for (size_t i = 0; i < num_samples; ++i) {
+        for (size_t j = 0; j < num_features; ++j) {
+            full_input_matrix(j, i) = x[j][i];
+        }
+    }
+
+    // Create indices for each cluster
+    std::vector<int> cluster_0_indices, cluster_1_indices;
+    for (size_t i = 0; i < cluster_labels.size(); ++i) {
+        if (cluster_labels[i] == 0) {
+            cluster_0_indices.push_back(i);
+        } else {
+            cluster_1_indices.push_back(i);
+        }
+    }
+
+    // Prepare matrices for each cluster
+    Eigen::MatrixXd input_matrix(num_features, cluster_0_indices.size());
+    Eigen::MatrixXd input_matrix_reactive(num_features, cluster_1_indices.size());
+
+    // Split data according to cluster labels
+    for (size_t i = 0; i < cluster_0_indices.size(); ++i) {
+        input_matrix.col(i) = full_input_matrix.col(cluster_0_indices[i]);
+    }
+    for (size_t i = 0; i < cluster_1_indices.size(); ++i) {
+        input_matrix_reactive.col(i) = full_input_matrix.col(cluster_1_indices[i]);
+    }
+    // Process each cluster
+    std::vector<double> result(num_samples * model.weight_matrices.back().rows());
+    Eigen_model_mutex->lock();
+
+    if (!cluster_0_indices.empty()) {
+        int num_batches_0 = std::ceil(static_cast<double>(cluster_0_indices.size()) / batch_size);
+        for (int batch = 0; batch < num_batches_0; ++batch) {
+            int start_idx = batch * batch_size;
+            int end_idx = std::min((batch + 1) * batch_size, static_cast<int>(cluster_0_indices.size()));
+            int current_batch_size = end_idx - start_idx;
+
+            Eigen::MatrixXd batch_data = input_matrix.block(0, start_idx, num_features, current_batch_size);
+            Eigen::MatrixXd batch_result = eigen_inference_batched(batch_data, model);
+
+            // Store results in their original positions
+            for (size_t i = 0; i < current_batch_size; ++i) {
+                int original_idx = cluster_0_indices[start_idx + i];
+                for (size_t j = 0; j < batch_result.rows(); ++j) {
+                    result[original_idx * batch_result.rows() + j] = batch_result(j, i);
+                }
+            }
+        }
+    }
+
+    // Process cluster 1
+    if (!cluster_1_indices.empty()) {
+        int num_batches_1 = std::ceil(static_cast<double>(cluster_1_indices.size()) / batch_size);
+        for (int batch = 0; batch < num_batches_1; ++batch) {
+            int start_idx = batch * batch_size;
+            int end_idx = std::min((batch + 1) * batch_size, static_cast<int>(cluster_1_indices.size()));
+            int current_batch_size = end_idx - start_idx;
+
+            Eigen::MatrixXd batch_data = input_matrix_reactive.block(0, start_idx, num_features, current_batch_size);
+            Eigen::MatrixXd batch_result = eigen_inference_batched(batch_data, model_reactive);
+
+            // Store results in their original positions
+            for (size_t i = 0; i < current_batch_size; ++i) {
+                int original_idx = cluster_1_indices[start_idx + i];
+                for (size_t j = 0; j < batch_result.rows(); ++j) {
+                    result[original_idx * batch_result.rows() + j] = batch_result(j, i);
+                }
+            }
+        }
+    }
+
+    Eigen_model_mutex->unlock();
+    return result;
+}
+
+/**
+ * @brief Uses the Eigen representation of the tKeras model weights for fast inference
+ * @param model The model weights and biases
+ * @param x 2D-Matrix with the content of a Field object
+ * @param batch_size size for mini-batches that are used in the Keras model.predict() method
+ * @param Eigen_model_mutex Mutex that locks the model during inference and prevents updaties from
+ * the training thread
+ * @return Predictions that the neural network made from the input values x. The predictions are
+ * represented as a vector similar to the representation from the Field.AsVector() method
+ */
+std::vector<double> Eigen_predict(const EigenModel& model, std::vector<std::vector<double>> x, int batch_size,
+                                  std::mutex* Eigen_model_mutex) {
+  // Convert input data to Eigen matrix
+  const int num_samples = x[0].size();
+  const int num_features = x.size();
+  Eigen::MatrixXd full_input_matrix(num_features, num_samples);
+
+  for (size_t i = 0; i < num_samples; ++i) {
+    for (size_t j = 0; j < num_features; ++j) {
+      full_input_matrix(j, i) = x[j][i];
+    }
+  }
+
+  std::vector<double> result;
+  result.reserve(num_samples * num_features);
+  
+  if (num_features != model.weight_matrices[0].cols()) {
+    throw std::runtime_error("Input data size " + std::to_string(num_features) + \
+      " does not match model input layer of size " + std::to_string(model.weight_matrices[0].cols()));
+  }
+  int num_batches = std::ceil(static_cast<double>(num_samples) / batch_size);
+
+  Eigen_model_mutex->lock();
+  for (int batch = 0; batch < num_batches; ++batch) {
+    int start_idx = batch * batch_size;
+    int end_idx = std::min((batch + 1) * batch_size, num_samples);
+    int current_batch_size = end_idx - start_idx;
+    // Extract the current input data batch
+    Eigen::MatrixXd batch_data(num_features, current_batch_size);
+    batch_data = full_input_matrix.block(0, start_idx, num_features, current_batch_size);
+    // Predict
+    batch_data = eigen_inference_batched(batch_data, model);
+
+    result.insert(result.end(), batch_data.data(), batch_data.data() + batch_data.size());
+  }
+  Eigen_model_mutex->unlock();
+  return result;
+}
+
+
+
+/**
+ * @brief Appends data from one matrix (column major std::vector<std::vector<double>>) to another
+ * @param training_data_buffer Matrix that the values are appended to
+ * @param new_values Matrix that is appended
+ */
+void training_data_buffer_append(std::vector<std::vector<double>>& training_data_buffer, 
+                                 std::vector<std::vector<double>>& new_values) {
+  // Initialize training data buffer if empty
+  if (training_data_buffer.size() == 0) {
+    training_data_buffer = new_values;
+  } else { // otherwise append
+    for (size_t col = 0; col < training_data_buffer.size(); col++) {
+      training_data_buffer[col].insert(training_data_buffer[col].end(),
+                                        new_values[col].begin(), new_values[col].end());
+    }
+  }
+}
+
+/**
+ * @brief Appends data from one int vector to another based on a mask vector
+ * @param labels Vector that the values are appended to
+ * @param new_labels Values that are appended
+ * @param validity Mask vector that defines how many and which values are appended
+ */
+void cluster_labels_append(std::vector<int>& labels_buffer, std::vector<int>& new_labels,
+                           std::vector<int> validity) {
+  // Calculate new buffer size from number of valid elements in mask
+  int n_invalid = validity.size();
+  for (size_t i = 0; i < validity.size(); i++) {
+    n_invalid -= validity[i];
+  }
+
+  // Resize label vector to hold non valid elements
+  int end_index = labels_buffer.size(); 
+  int new_size = end_index + n_invalid;
+  labels_buffer.resize(new_size);
+  // Iterate over mask to transfer cluster labels
+  for (size_t i = 0; i < validity.size(); ++i) {
+    // Append only the labels of invalid rows 
+    if (!validity[i]) {
+      labels_buffer[end_index] = new_labels[i]; 
+      end_index++;
+    }
+  }
+}
+
+
+/**
+ * @brief Uses the Python environment with Keras' default functions to train the model
+ * @param x Training data features
+ * @param y Training data targets
+ * @param params Global runtime paramters
+ */
+void Python_Keras_train(std::vector<std::vector<double>>& x, std::vector<std::vector<double>>& y,
+                        int train_cluster, std::string model_name, const RuntimeParameters& params) {
+  // Prepare data for python
+  PyObject* py_df_x = vector_to_numpy_array(x);
+  PyObject* py_df_y = vector_to_numpy_array(y);
+
+  // Make sure that model output file name .keras file
+  std::string model_path = params.save_model_path; 
+  if (!model_path.empty()) {
+    if (model_path.length() >= 6 && model_path.substr(model_path.length() - 6) != ".keras") {
+        model_path += ".keras";
+    }
+  }
+
+  // Choose the correct model to train if clustering is used
+  if (train_cluster == 1) {
+    if (!model_path.empty()) {
+      model_path.insert(model_path.length() - 6, "_reaction");
+      std::cout << "MODEL SAVED AS:" << model_path << std::endl;
+    }
+  }
+
+  // Get the model and training function from the global python interpreter
+  PyObject* py_main_module = PyImport_AddModule("__main__");
+  PyObject* py_global_dict = PyModule_GetDict(py_main_module);
+  PyObject* py_keras_model = PyDict_GetItemString(py_global_dict, model_name.c_str());
+  PyObject* py_training_function = PyDict_GetItemString(py_global_dict, "training_step");
+  
+  // Build the function arguments as four python objects and an integer
+  PyObject* args = Py_BuildValue("(OOOiis)", 
+      py_keras_model, py_df_x, py_df_y, params.batch_size, params.training_epochs,
+      model_path.c_str());
+
+  // Call the Python training function
+  PyObject *py_rv = PyObject_CallObject(py_training_function, args);
+
+  // Clean up
+  PyErr_Print(); 
+  Py_DECREF(py_df_x);
+  Py_DECREF(py_df_y);
+  Py_DECREF(args);
+}
+
+/**
+ * @brief Function for threadsafe parallel training and weight updating.
+ * The function waits conditionally until the training data buffer is full.
+ * It then clears the buffer and starts training, after training it writes the new weights to 
+ * the Eigen model.  
+ * @param Eigen_model Pointer to the EigenModel struct that will be updates with new weights
+ * @param Eigen_model_mutex Mutex to ensure threadsafe access to the EigenModel struct
+ * @param training_data_buffer Pointer to the Training data struct with which the model is trained
+ * @param training_data_buffer_mutex Mutex to ensure threadsafe access to the training data struct
+ * @param training_data_buffer_full Conditional waiting variable with wich the main thread signals
+ * when a training run can start
+ * @param start_training Conditional waiting predicate to mitigate against spurious wakeups
+ * @param end_training Signals end of program to wind down thread gracefully 
+ * @param params Global runtime paramters
+ * @return 0 if function was succesful
+ */
+void parallel_training(EigenModel* Eigen_model, EigenModel* Eigen_model_reactive,
+                       std::mutex* Eigen_model_mutex,
+                       TrainingData* training_data_buffer,
+                       std::mutex* training_data_buffer_mutex,
+                       std::condition_variable* training_data_buffer_full,
+                       bool* start_training, bool* end_training,
+                       const RuntimeParameters& params) {
+  while (true) {
+    // Conditional waiting:
+    // - Sleeps until a signal arrives on training_data_buffer_full
+    // - Releases the lock on training_data_buffer_mutex while sleeping
+    // - Lambda function with start_training checks if it was a spurious wakeup
+    // - Reaquires the lock on training_data_buffer_mutex after waking up
+    // - If start_training has been set to true while the thread was active, it does NOT
+    //   wait for a signal on training_data_buffer_full but starts the next round immediately.
+    std::unique_lock<std::mutex> lock(*training_data_buffer_mutex);
+    training_data_buffer_full->wait(lock, [start_training] { return *start_training;});
+    // Return if program is about to end
+    if (*end_training) {
+      return;
+    }
+    // Get the necessary training data
+    std::cout << "AI: Training thread: Getting training data" << std::endl;
+    // Initialize training data input and targets
+    std::vector<std::vector<double>> inputs(training_data_buffer->x.size(),
+                                            std::vector<double>(params.training_data_size));
+    std::vector<std::vector<double>> targets(training_data_buffer->y.size(),
+                                             std::vector<double>(params.training_data_size));
+
+    int buffer_size = training_data_buffer->x[0].size();
+    // If clustering is used, check the current cluster
+    int n_cluster_reactive = 0;
+    int train_cluster = -1; // Default value for non clustered training (all data is used)
+    if (params.use_k_means_clustering) {
+      for (size_t i = 0; i < buffer_size; i++) {
+          n_cluster_reactive += training_data_buffer->cluster_labels[i];
+      }
+      train_cluster = n_cluster_reactive >= params.training_data_size;
+    }
+    int buffer_row = 0;
+    int copied_row = 0;
+    while (copied_row < params.training_data_size) {
+      if ((train_cluster == -1) ||
+          (train_cluster == training_data_buffer->cluster_labels[buffer_row])) {
+        for (size_t col = 0; col < training_data_buffer->x.size(); col++) {
+          // Copy and remove from training data buffer
+          inputs[col][copied_row] = training_data_buffer->x[col][buffer_row];
+          targets[col][copied_row] = training_data_buffer->y[col][buffer_row];
+          training_data_buffer->x[col].erase(training_data_buffer->x[col].begin() + 
+                                            buffer_row);
+          training_data_buffer->y[col].erase(training_data_buffer->y[col].begin() +
+                                            buffer_row);
+        }
+        // Remove from cluster label buffer
+        if (params.use_k_means_clustering) {
+          training_data_buffer->cluster_labels.erase(
+            training_data_buffer->cluster_labels.begin() + buffer_row);
+        }
+        copied_row++;
+      } else {
+        buffer_row++;
+      }
+    }
+
+    // Set the waiting predicate to immediately continue training if enough elements of any cluster remain
+    if (train_cluster == 1) {
+      *start_training = ((n_cluster_reactive - params.training_data_size) >= params.training_data_size) ||
+                        ((buffer_size - n_cluster_reactive) >= params.training_data_size);
+    } else {
+      *start_training = (buffer_size - n_cluster_reactive - params.training_data_size)
+                        >= params.training_data_size;
+    }
+    
+    //update number of training runs
+    training_data_buffer->n_training_runs += 1;
+    // Unlock the training_data_buffer_mutex 
+    lock.unlock();
+
+    std::string model_name = "model";
+    if (train_cluster == 1) {
+      model_name = "model_reactive";
+    }
+    std::cout << "AI: Training thread: Start training " << model_name << std::endl;
+
+    // Acquire the Python GIL
+    PyGILState_STATE gstate = PyGILState_Ensure();
+    // Start training
+    Python_Keras_train(inputs, targets, train_cluster, model_name, params);
+    
+    if (!params.use_Keras_predictions) {
+      std::cout << "AI: Training thread: Update shared model weights" << std::endl;
+      std::vector<std::vector<std::vector<double>>> cpp_weights = Python_Keras_get_weights(model_name);
+      Eigen_model_mutex->lock();
+      if (train_cluster == 1) {
+        update_weights(Eigen_model_reactive, cpp_weights);
+      } else {
+        update_weights(Eigen_model, cpp_weights);
+      }
+      Eigen_model_mutex->unlock();
+    }
+    
+    // Release the Python GIL
+    PyGILState_Release(gstate);
+    std::cout << "AI: Training thread: Finished training, waiting for new data" << std::endl;
+  }
+}
+
+std::thread python_train_thread;
+/**
+ * @brief Starts a thread for parallel training and weight updating. This Wrapper function
+ * ensures, that the main POET program can be built without pthread support if the AI
+ * surrogate functions are disabled during compilation. 
+ * @param Eigen_model Pointer to the EigenModel struct that will be updates with new weights
+ * @param Eigen_model_mutex Mutex to ensure threadsafe access to the EigenModel struct
+ * @param training_data_buffer Pointer to the Training data struct with which the model is trained
+ * @param training_data_buffer_mutex Mutex to ensure threadsafe access to the training data struct
+ * @param training_data_buffer_full Conditional waiting variable with wich the main thread signals
+ * when a training run can start
+ * @param start_training Conditional waiting predicate to mitigate against spurious wakeups
+ * @param end_training Signals end of program to wind down thread gracefully 
+ * @param params Global runtime paramters
+ * @return 0 if function was succesful
+ */
+int Python_Keras_training_thread(EigenModel* Eigen_model, EigenModel* Eigen_model_reactive,
+                                 std::mutex* Eigen_model_mutex,
+                                 TrainingData* training_data_buffer,
+                                 std::mutex* training_data_buffer_mutex,
+                                 std::condition_variable* training_data_buffer_full,
+                                 bool* start_training, bool* end_training,
+                                 const RuntimeParameters& params) {
+  PyThreadState *_save = PyEval_SaveThread();
+  python_train_thread = std::thread(parallel_training, Eigen_model, Eigen_model_reactive,
+                                    Eigen_model_mutex, training_data_buffer,
+                                    training_data_buffer_mutex, training_data_buffer_full,
+                                    start_training, end_training, params);
+  return 0;
+}
+
+
+/**
+ * @brief Updates the EigenModels weigths and biases from the weight vector
+ * @param model Pounter to an EigenModel struct
+ * @param weights Cector of model weights from keras as returned by Python_Keras_get_weights()
+ */
+void update_weights(EigenModel* model,
+                    const std::vector<std::vector<std::vector<double>>>& weights) {
+  size_t num_layers = weights.size() / 2;
+  for (size_t i = 0; i < weights.size(); i += 2) {
+      // Fill current weight matrix
+      size_t rows = weights[i][0].size();
+      size_t cols = weights[i].size();
+      for (size_t j = 0; j < cols; ++j) {
+          for (size_t k = 0; k < rows; ++k) {
+              model->weight_matrices[i / 2](k, j) = weights[i][j][k];
+          }
+      }
+      // Fill bias vector
+      size_t bias_size = weights[i + 1][0].size();
+      for (size_t j = 0; j < bias_size; ++j) {
+          model->biases[i / 2](j) = weights[i + 1][0][j];
+      }
+  }
+}
+
+
+/**
+ * @brief Converts the weights and biases from the Python Keras model to C++ vectors
+ * @return A vector containing the model weights and biases 
+ */
+std::vector<std::vector<std::vector<double>>> Python_Keras_get_weights(std::string model_name) {
+  // Acquire the Python GIL
+  PyGILState_STATE gstate = PyGILState_Ensure();
+
+  PyObject* py_main_module = PyImport_AddModule("__main__");
+  PyObject* py_global_dict = PyModule_GetDict(py_main_module);
+  PyObject* py_keras_model = PyDict_GetItemString(py_global_dict, model_name.c_str());
+  PyObject* py_get_weights_function = PyDict_GetItemString(py_global_dict, "get_weights");
+  PyObject* args = Py_BuildValue("(O)", py_keras_model);
+  
+  // Call Python function
+  PyObject* py_weights_list = PyObject_CallObject(py_get_weights_function, args);
+  
+  if (!py_weights_list) {
+      PyErr_Print();
+      throw std::runtime_error("Failed to get weights from Keras model");
+  }
+
+  // Container for the extracted weights
+  std::vector<std::vector<std::vector<double>>> cpp_weights;
+
+  // Iterate through the layers (weights and biases)
+  Py_ssize_t num_layers = PyList_Size(py_weights_list);
+  for (Py_ssize_t i = 0; i < num_layers; ++i) {
+    PyObject* py_weight_array = PyList_GetItem(py_weights_list, i);
+    if (!PyArray_Check(py_weight_array)) {
+      throw std::runtime_error("Weight is not a NumPy array.");
+    }
+
+    PyArrayObject* weight_np = reinterpret_cast<PyArrayObject*>(py_weight_array);
+    int dtype = PyArray_TYPE(weight_np);
+
+    // If array is 2D it's a weight matrix
+    if (PyArray_NDIM(weight_np) == 2) {
+      int num_rows = PyArray_DIM(weight_np, 0);
+      int num_cols = PyArray_DIM(weight_np, 1);
+      
+      std::vector<std::vector<double>> weight_matrix(num_rows, std::vector<double>(num_cols));
+      // Handle different precision settings
+      if (dtype == NPY_FLOAT32) {
+        float* weight_data_float = static_cast<float*>(PyArray_DATA(weight_np));
+        for (size_t r = 0; r < num_rows; ++r) {
+          for (size_t c = 0; c < num_cols; ++c) {
+            weight_matrix[r][c] = static_cast<double>(weight_data_float[r * num_cols + c]);
+          }
+        }
+      } else if (dtype == NPY_DOUBLE) {
+        double* weight_data_double = static_cast<double*>(PyArray_DATA(weight_np));
+        for (size_t r = 0; r < num_rows; ++r) {
+          for (size_t c = 0; c < num_cols; ++c) {
+            weight_matrix[r][c] = weight_data_double[r * num_cols + c];
+          }
+        }
+      } else {
+          throw std::runtime_error("Unsupported data type for weights. Must be NPY_FLOAT32 or NPY_DOUBLE.");
+      }
+      cpp_weights.push_back(weight_matrix);
+
+    // If array is 1D it's a bias vector 
+    } else if (PyArray_NDIM(weight_np) == 1) {
+      int num_elements = PyArray_DIM(weight_np, 0);
+      std::vector<std::vector<double>> bias_vector(1, std::vector<double>(num_elements));
+
+      // Handle different precision settings
+      if (dtype == NPY_FLOAT32) {
+        float* bias_data_float = static_cast<float*>(PyArray_DATA(weight_np));
+        for (size_t j = 0; j < num_elements; ++j) {
+          bias_vector[0][j] = static_cast<double>(bias_data_float[j]);
+        }
+      } else if (dtype == NPY_DOUBLE) {
+        double* bias_data_double = static_cast<double*>(PyArray_DATA(weight_np));
+        for (size_t j = 0; j < num_elements; ++j) {
+          bias_vector[0][j] = bias_data_double[j];
+        }
+      } else {
+        throw std::runtime_error("Unsupported data type for biases. Must be NPY_FLOAT32 or NPY_DOUBLE.");
+      }
+      cpp_weights.push_back(bias_vector);
+    }
+  }
+  // Clean up
+  Py_DECREF(py_weights_list);
+  Py_DECREF(args);
+  // Release Python GIL
+  PyGILState_Release(gstate);
+
+  return cpp_weights;
+}
+
+
+/**
+ * @brief Joins the training thread and winds down the Python environment gracefully
+ * @param Eigen_model_mutex Mutex to ensure threadsafe access to the EigenModel struct
+ * @param training_data_buffer_mutex Mutex to ensure threadsafe access to the training data struct
+ * @param training_data_buffer_full Conditional waiting variable with wich the main thread signals
+ * when a training run can start
+ * @param start_training Conditional waiting predicate to mitigate against spurious wakeups
+ * @param end_training Signals end of program to wind down thread gracefully */
+void Python_finalize(std::mutex* Eigen_model_mutex, std::mutex* training_data_buffer_mutex,
+                     std::condition_variable* training_data_buffer_full,
+                     bool* start_training, bool* end_training) {
+  training_data_buffer_mutex->lock();
+  // Define training as over
+  *end_training = true;
+  // Wake up and join training thread
+  *start_training = true;
+  training_data_buffer_mutex->unlock();
+  training_data_buffer_full->notify_one();
+  
+  if (python_train_thread.joinable()) {
+    python_train_thread.join();
+  }
+
+  // Acquire the Python GIL
+  PyGILState_STATE gstate = PyGILState_Ensure();
+  //Finalize Python
+  Py_FinalizeEx();
+}
+
+} //namespace poet

src/Chemistry/SurrogateModels/AI_functions.hpp

@@ -0,0 +1,120 @@
+/**
+ * @file AI_functions.hpp
+ * @author Hans Straile (straile@uni-potsdam.de)
+ * @brief API for the AI/Machine Learning based chemistry surrogate model with functions to initialize a neural network and use it for training and inference via Keras for Python . 
+ * @version 0.1
+ * @date 01 Nov 2024
+ *
+ * This file implements functions to train and predict with a neural network.
+ * All functions are based on a user supplied Keras model. Pyhton.h is used for model
+ * training with Keras and can be used for inference. The default inference funtion 
+ * is implemented with Eigen matrices in C++. All functions use 2 different models
+ * to process data separately according to a K-means cluster assignement. This file
+ * alo contains the functions for the K-means algorithm.
+ *  
+ */
+
+#ifndef AI_FUNCTIONS_H
+#define AI_FUNCTIONS_H
+
+#include <string>
+#include <vector>
+#include "poet.hpp"
+
+// PhreeqC definition of pi clashes with Eigen macros
+// so we have to temporarily undef it 
+#pragma push_macro("pi")
+#undef pi
+#include <Eigen/Dense> 
+#pragma pop_macro("pi")
+
+namespace poet {
+
+struct EigenModel {
+  // The first model will be used for all values if clustering is disabled
+  // or for the reactive part of the field if clustering is enabled
+  std::vector<Eigen::MatrixXd> weight_matrices;
+  std::vector<Eigen::VectorXd> biases;
+
+  // The other model will be used for the non-reactive cluster
+  // (if clustering is enabled)
+  std::vector<Eigen::MatrixXd> weight_matrices_no_reaction;
+  std::vector<Eigen::VectorXd> biases_no_reaction;
+};
+
+struct TrainingData {
+  std::vector<std::vector<double>> x;
+  std::vector<std::vector<double>> y;
+  std::vector<int> cluster_labels;
+  int n_training_runs = 0;
+};
+
+// Ony declare the actual functions if flag is set 
+#ifdef USE_AI_SURROGATE
+
+int Python_Keras_setup(std::string functions_file_path, std::string cuda_src_dir);
+
+void Python_finalize(std::mutex* Eigen_model_mutex, std::mutex* training_data_buffer_mutex,
+                     std::condition_variable* training_data_buffer_full, bool* start_training, 
+                     bool* end_training);
+
+int Python_Keras_load_model(std::string model, std::string model_reactive,
+                            bool use_clustering);
+
+std::vector<int> K_Means(std::vector<std::vector<double>>& field, int k, int maxIterations = 100);
+
+std::vector<double> Python_Keras_predict(std::vector<std::vector<double>>& x, int batch_size,
+                                         std::vector<int>& cluster_labels);  
+
+void training_data_buffer_append(std::vector<std::vector<double>>& training_data_buffer,
+                                 std::vector<std::vector<double>>& new_values);
+
+void cluster_labels_append(std::vector<int>& labels, std::vector<int>& new_labels,
+                           std::vector<int> validity);
+
+int Python_Keras_training_thread(EigenModel* Eigen_model, EigenModel* Eigen_model_reactive,
+                                 std::mutex* Eigen_model_mutex,
+                                 TrainingData* training_data_buffer,
+                                 std::mutex* training_data_buffer_mutex,
+                                 std::condition_variable* training_data_buffer_full,
+                                 bool* start_training, bool* end_training,
+                                 const RuntimeParameters& params);
+
+void update_weights(EigenModel* model, const std::vector<std::vector<std::vector<double>>>& weights);
+
+std::vector<std::vector<std::vector<double>>> Python_Keras_get_weights(std::string model_name);
+
+std::vector<double> Eigen_predict_clustered(const EigenModel& model, const EigenModel& model_reactive,
+                                            std::vector<std::vector<double>>& x, 
+                                            int batch_size, std::mutex* Eigen_model_mutex,
+                                            std::vector<int>& cluster_labels);
+std::vector<double> Eigen_predict(const EigenModel& model, std::vector<std::vector<double>> x, int batch_size,
+                                  std::mutex* Eigen_model_mutex);
+
+
+// Otherwise, define the necessary stubs
+#else
+inline void Python_Keras_setup(std::string, std::string){}
+inline void Python_finalize(std::mutex*, std::mutex*, std::condition_variable*, bool*, bool*){}
+inline void Python_Keras_load_model(std::string, std::string, bool){}
+inline std::vector<int> K_Means(std::vector<std::vector<double>>&, int, int) {return {};}
+inline std::vector<double> Python_Keras_predict(std::vector<std::vector<double>>&, int,
+                                                std::vector<int>&){return {};}
+inline void training_data_buffer_append(std::vector<std::vector<double>>&,
+                                        std::vector<std::vector<double>>&){}
+inline void cluster_labels_append(std::vector<int>&, std::vector<int>&, std::vector<int>){}
+inline int Python_Keras_training_thread(EigenModel*, EigenModel*, std::mutex*, 
+                                        TrainingData*, std::mutex*, std::condition_variable*,
+                                        bool*, bool*, const RuntimeParameters&){return {};}
+
+inline void update_weights(EigenModel*, const std::vector<std::vector<std::vector<double>>>&){}
+inline std::vector<std::vector<std::vector<double>>> Python_Keras_get_weights(std::string){return {};}
+inline std::vector<double> Eigen_predict_clustered(const EigenModel&, const EigenModel&,
+                                                   std::vector<std::vector<double>>&, int,
+                                                   std::mutex*, std::vector<int>&){return {};}
+inline std::vector<double> Eigen_predict(const EigenModel&, std::vector<std::vector<double>>, int,
+                                         std::mutex*){return {};}
+#endif
+} // namespace poet
+
+#endif // AI_FUNCTIONS_HPP

src/Chemistry/WorkerFunctions.cpp

@@ -161,7 +161,15 @@ void poet::ChemistryModule::WorkerDoWork(MPI_Status &probe_status,
                             mpi_buffer.begin() + this->prop_count * (wp_i + 1));
   }
 
-  // std::cout << this->comm_rank << ":" << counter++ << std::endl;
+  if (this->ai_surrogate_enabled) {
+    // Map valid predictions from the ai surrogate in the workpackage
+    for (int i = 0; i < s_curr_wp.size; i++) {
+      if (this->ai_surrogate_validity_vector[wp_start_index + i] == 1) {
+        s_curr_wp.mapping[i] = CHEM_AISURR;
+      }
+    }
+  }
+
   if (dht_enabled || interp_enabled) {
     dht->prepareKeys(s_curr_wp.input, dt);
   }
@@ -178,16 +186,6 @@ void poet::ChemistryModule::WorkerDoWork(MPI_Status &probe_status,
     interp->tryInterpolation(s_curr_wp);
   }
 
-  if (this->ai_surrogate_enabled) {
-    // Map valid predictions from the ai surrogate in the workpackage
-    for (int i = 0; i < s_curr_wp.size; i++) {
-      if (this->ai_surrogate_validity_vector[wp_start_index + i] == 1) {
-        s_curr_wp.mapping[i] = CHEM_AISURR;
-      }
-    }
-  }
-
-
   phreeqc_time_start = MPI_Wtime();
 
   WorkerRunWorkPackage(s_curr_wp, current_sim_time, dt);

src/Init/ChemistryInit.cpp

@@ -51,7 +51,7 @@ void InitialList::initChemistry(const Rcpp::List &chem) {
     // Get base path
     ai_surrogate_input_script_path = ai_surrogate_input_script_path.substr(0, ai_surrogate_input_script_path.find_last_of('/') + 1);
     // Add the filepath as a global variable in R to enable relative filepaths in the R script
-    fileContent += "\nai_surrogate_base_path <- \"" + ai_surrogate_input_script_path + "\"";
+    fileContent.insert(0, "ai_surrogate_base_path <- \"" + ai_surrogate_input_script_path + "\"\n");
 
     this->ai_surrogate_input_script = fileContent;
   }

src/poet.cpp

@@ -28,7 +28,6 @@
 #include "DataStructures/Field.hpp"
 #include "Init/InitialList.hpp"
 #include "Transport/DiffusionModule.hpp"
-
 #include <RInside.h>
 #include <Rcpp.h>
 #include <Rcpp/DataFrame.h>
@@ -39,9 +38,10 @@
 #include <memory>
 #include <mpi.h>
 #include <string>
-
+#include <mutex>
+#include <condition_variable>
+#include "Chemistry/SurrogateModels/AI_functions.hpp"
 #include <CLI/CLI.hpp>
-
 #include <poet.hpp>
 #include <vector>
 
@@ -72,6 +72,7 @@ static void init_global_functions(RInside &R) {
   SaveRObj_R = DEFunc("SaveRObj");
 }
 
+
 // HACK: this is a step back as the order and also the count of fields is
 // predefined, but it will change in the future
 // static inline void writeFieldsToR(RInside &R, const Field &trans,
@@ -182,6 +183,12 @@ int parseInitValues(int argc, char **argv, RuntimeParameters &params) {
     MSG("Work Package Size: " + std::to_string(params.work_package_size));
     MSG("DHT is " + BOOL_PRINT(params.use_dht));
     MSG("AI Surrogate is " + BOOL_PRINT(params.use_ai_surrogate));
+    #ifndef USE_AI_SURROGATE
+      if (params.use_ai_surrogate) {
+        throw std::runtime_error("AI Surrogate functions can only be used if they are included during compile time.\n \
+          Please use the CMake flag -DUSE_AI_SURROGATE=ON.");
+      }
+    #endif
 
     if (params.use_dht) {
       // MSG("DHT strategy is " + std::to_string(simparams.dht_strategy));
@@ -278,7 +285,77 @@ static Rcpp::List RunMasterLoop(RInsidePOET &R, const RuntimeParameters &params,
   if (params.print_progress) {
     chem.setProgressBarPrintout(true);
   } 
+
+  /* For the weights and biases of the AI surrogate
+   * model to use in an inference function with Eigen */
+  std::mutex Eigen_model_mutex; 
+  static EigenModel Eigen_model;
+  
+  /* With K Means clustering, a second model will be used 
+   * only for the reactive part of the field (and the first
+   * model only for the non-reactive part) */
+  static EigenModel Eigen_model_reactive;
+  
+
+  /* For the training data */
+  std::mutex training_data_buffer_mutex;
+  std::condition_variable training_data_buffer_full;
+  bool start_training, end_training;
+  TrainingData training_data_buffer;
+  std::vector<int> cluster_labels;
+  if (params.use_ai_surrogate) {  
+    MSG("AI: Initialize model");
+
+    // Initiate two models from one file
+    Python_Keras_load_model(R["model_file_path"], R["model_reactive_file_path"],
+                            params.use_k_means_clustering);
+    if (!params.disable_training) {
+      MSG("AI: Initialize training thread");
+      Python_Keras_training_thread(&Eigen_model, &Eigen_model_reactive, 
+                                   &Eigen_model_mutex, &training_data_buffer, 
+                                   &training_data_buffer_mutex,
+                                   &training_data_buffer_full, &start_training,
+                                   &end_training, params);
+    }
+    if (!params.use_Keras_predictions) {
+      // Initialize Eigen model for custom inference function
+      MSG("AI: Use custom C++ prediction function");
+      // Get Keras weights from Python
+      std::vector<std::vector<std::vector<double>>> cpp_weights = Python_Keras_get_weights("model");
+      // Set model size
+      size_t num_layers = cpp_weights.size() / 2;
+      Eigen_model.weight_matrices.resize(num_layers); 
+      Eigen_model.biases.resize(num_layers);
+      for (size_t i = 0; i < cpp_weights.size(); i += 2) {
+          size_t rows = cpp_weights[i][0].size();
+          size_t cols = cpp_weights[i].size();
+          Eigen_model.weight_matrices[i / 2].resize(rows, cols);
+          size_t bias_size = cpp_weights[i + 1][0].size();
+          Eigen_model.biases[i / 2].resize(bias_size);
+      }
+      // Set initial model weights
+      update_weights(&Eigen_model, cpp_weights);
+      if (params.use_k_means_clustering) {
+        // Initialize Eigen model for reactive part of the field
+        cpp_weights = Python_Keras_get_weights("model_reactive");
+        num_layers = cpp_weights.size() / 2;
+        Eigen_model_reactive.weight_matrices.resize(num_layers); 
+        Eigen_model_reactive.biases.resize(num_layers);
+        for (size_t i = 0; i < cpp_weights.size(); i += 2) {
+            size_t rows = cpp_weights[i][0].size();
+            size_t cols = cpp_weights[i].size();
+            Eigen_model_reactive.weight_matrices[i / 2].resize(rows, cols);
+            size_t bias_size = cpp_weights[i + 1][0].size();
+            Eigen_model_reactive.biases[i / 2].resize(bias_size);
+        }
+        update_weights(&Eigen_model_reactive, cpp_weights);
+      }
+    }
+    MSG("AI: Surrogate model initialized");
+  }
+
   R["TMP_PROPS"] = Rcpp::wrap(chem.getField().GetProps());
+  R["field_nrow"] = chem.getField().GetRequestedVecSize();
 
     /* SIMULATION LOOP */
   double dSimTime{0};
@@ -299,77 +376,72 @@ static Rcpp::List RunMasterLoop(RInsidePOET &R, const RuntimeParameters &params,
     chem.getField().update(diffusion.getField());
 
     MSG("Chemistry step");
+    
     if (params.use_ai_surrogate) {
       double ai_start_t = MPI_Wtime();
-      // Save current values from the tug field as predictor for the ai step
+
+      // Get current values from the tug field for the ai predictions
       R["TMP"] = Rcpp::wrap(chem.getField().AsVector());
-      R.parseEval(
-          std::string("predictors <- setNames(data.frame(matrix(TMP, nrow=" +
-                      std::to_string(chem.getField().GetRequestedVecSize()) +
-                      ")), TMP_PROPS)"));
-      R.parseEval("predictors <- predictors[ai_surrogate_species]");
+      R.parseEval(std::string("predictors <- ") + 
+        "set_field(TMP, TMP_PROPS, field_nrow, ai_surrogate_species)");
 
       // Apply preprocessing
       MSG("AI Preprocessing");
       R.parseEval("predictors_scaled <- preprocess(predictors)");
+      std::vector<std::vector<double>> predictors_scaled = R["predictors_scaled"];
 
-      // Predict
-      MSG("AI Predict");
-      R.parseEval(
-          "aipreds_scaled <- prediction_step(model, predictors_scaled)");
+      // Get K-Means cluster assignements based on the preprocessed data
+      if (params.use_k_means_clustering) {
+        cluster_labels = K_Means(predictors_scaled, 2, 300);
+        R["cluster_labels"] = cluster_labels;
+      }
+      
+      MSG("AI: Predict");
+      if (params.use_Keras_predictions) {  // Predict with Keras default function
+        R["TMP"] = Python_Keras_predict(predictors_scaled, params.batch_size, cluster_labels);
+      } else {  // Predict with custom Eigen function
+        if (params.use_k_means_clustering) {
+          R["TMP"] = Eigen_predict_clustered(Eigen_model, Eigen_model_reactive, 
+                                             predictors_scaled, params.batch_size,
+                                             &Eigen_model_mutex, cluster_labels);
+        } else {
+          R["TMP"] = Eigen_predict(Eigen_model, predictors_scaled, params.batch_size, 
+                                   &Eigen_model_mutex);
+        }
+      }
 
       // Apply postprocessing
-      MSG("AI Postprocesing");
-      R.parseEval("aipreds <- postprocess(aipreds_scaled)");
+      MSG("AI: Postprocesing");
+      R.parseEval(std::string("predictions_scaled <- ") + 
+        "set_field(TMP, ai_surrogate_species, field_nrow, ai_surrogate_species, byrow = TRUE)");
+      R.parseEval("predictions <- postprocess(predictions_scaled)");
 
       // Validate prediction and write valid predictions to chem field
-      MSG("AI Validate");
-      R.parseEval(
-          "validity_vector <- validate_predictions(predictors, aipreds)");
+      MSG("AI: Validate");
+      R.parseEval("validity_vector <- validate_predictions(predictors, predictions)");
 
-      MSG("AI Marking accepted");
+      MSG("AI: Marking valid");
       chem.set_ai_surrogate_validity_vector(R.parseEval("validity_vector"));
 
-      MSG("AI TempField");
       std::vector<std::vector<double>> RTempField =
-          R.parseEval("set_valid_predictions(predictors,\
-                                                                                       aipreds,\
-                                                                                       validity_vector)");
+        R.parseEval("set_valid_predictions(predictors, predictions, validity_vector)");
 
-      MSG("AI Set Field");
       Field predictions_field =
-          Field(R.parseEval("nrow(predictors)"), RTempField,
-                R.parseEval("colnames(predictors)"));
+          Field(R.parseEval("nrow(predictions)"), RTempField,
+                R.parseEval("colnames(predictions)"));
 
-      MSG("AI Update");
+      MSG("AI: Update field with AI predictions");
       chem.getField().update(predictions_field);
+      
+      // store time for output file
       double ai_end_t = MPI_Wtime();
       R["ai_prediction_time"] = ai_end_t - ai_start_t;
     }
 
+    // Run simulation step
+    MSG("Simulate chemistry");
     chem.simulate(dt);
 
-    /* AI surrogate iterative training*/
-    if (params.use_ai_surrogate) {
-      double ai_start_t = MPI_Wtime();
-
-      R["TMP"] = Rcpp::wrap(chem.getField().AsVector());
-      R.parseEval(
-          std::string("targets <- setNames(data.frame(matrix(TMP, nrow=" +
-                      std::to_string(chem.getField().GetRequestedVecSize()) +
-                      ")), TMP_PROPS)"));
-      R.parseEval("targets <- targets[ai_surrogate_species]");
-
-      // TODO: Check how to get the correct columns
-      R.parseEval("target_scaled <- preprocess(targets)");
-
-      MSG("AI: incremental training");
-      R.parseEval("model <- training_step(model, predictors_scaled, "
-                  "target_scaled, validity_vector)");
-      double ai_end_t = MPI_Wtime();
-      R["ai_training_time"] = ai_end_t - ai_start_t;
-    }
-
     // MPI_Barrier(MPI_COMM_WORLD);
     double end_t = MPI_Wtime();
     dSimTime += end_t - start_t;
@@ -381,6 +453,44 @@ static Rcpp::List RunMasterLoop(RInsidePOET &R, const RuntimeParameters &params,
     // store_result is TRUE)
     call_master_iter_end(R, diffusion.getField(), chem.getField());
 
+    /* AI surrogate iterative training*/
+    if (params.use_ai_surrogate && !params.disable_training) {
+      MSG("AI: Add to training data buffer");
+      if (!params.train_only_invalid) {
+        // Use all values if not specified otherwise 
+        R.parseEval("validity_vector <- rep(0, length(validity_vector))");
+      }
+      std::vector<std::vector<double>> invalid_x = 
+        R.parseEval("get_invalid_values(predictors_scaled, validity_vector)");
+
+      R.parseEval("target_scaled <- preprocess(state_C[ai_surrogate_species])");
+      std::vector<std::vector<double>> invalid_y = 
+        R.parseEval("get_invalid_values(target_scaled, validity_vector)");
+
+      training_data_buffer_mutex.lock();
+      training_data_buffer_append(training_data_buffer.x, invalid_x);
+      training_data_buffer_append(training_data_buffer.y, invalid_y);
+            
+      // If clustering is used, Add cluster labels to buffer and
+      // count buffer size according to the cluster assignements
+      int n_cluster_reactive = 0;
+      size_t  buffer_size = training_data_buffer.x[0].size();
+      if (params.use_k_means_clustering) {
+        cluster_labels_append(training_data_buffer.cluster_labels, cluster_labels,
+                              R["validity_vector"]);
+        for (size_t i = 0; i < buffer_size; i++) {
+          n_cluster_reactive += training_data_buffer.cluster_labels[i];
+        }
+      }
+      // Signal to the training thread if enough training data is present
+      if ((n_cluster_reactive >= params.training_data_size) ||
+         (buffer_size - n_cluster_reactive >= params.training_data_size)) {
+        start_training = true;
+        training_data_buffer_full.notify_one();
+      }
+      training_data_buffer_mutex.unlock();
+    }
+
     diffusion.getField().update(chem.getField());
 
     MSG("End of *coupling* iteration " + std::to_string(iter) + "/" +
@@ -426,6 +536,12 @@ static Rcpp::List RunMasterLoop(RInsidePOET &R, const RuntimeParameters &params,
   profiling["chemistry"] = chem_profiling;
   profiling["diffusion"] = diffusion_profiling;
 
+  if (params.use_ai_surrogate) {
+    MSG("Finalize Python and wind down training thread");
+    Python_finalize(&Eigen_model_mutex, &training_data_buffer_mutex,
+                    &training_data_buffer_full, &start_training, &end_training);
+  }
+
   chem.MasterLoopBreak();
 
   return profiling;
@@ -477,7 +593,11 @@ std::vector<std::string> getSpeciesNames(const Field &&field, int root,
 int main(int argc, char *argv[]) {
   int world_size;
 
-  MPI_Init(&argc, &argv);
+  // Threadsafe MPI is necessary for the AI surrogate
+  // training thread
+  int provided;
+  int required = MPI_THREAD_FUNNELED;
+  MPI_Init_thread(&argc, &argv, required, &provided);
 
   {
     MPI_Comm_size(MPI_COMM_WORLD, &world_size);
@@ -548,8 +668,15 @@ int main(int argc, char *argv[]) {
       R["out_ext"] = run_params.out_ext;
       R["out_dir"] = run_params.out_dir;
 
+      // MPI_Barrier(MPI_COMM_WORLD);
+
+      DiffusionModule diffusion(init_list.getDiffusionInit(),
+                                init_list.getInitialGrid());
+
+      chemistry.masterSetField(init_list.getInitialGrid());
+
       if (run_params.use_ai_surrogate) {
-        /* Incorporate ai surrogate from R */
+        // Load default function implementations 
         R.parseEvalQ(ai_surrogate_r_library);
         /* Use dht species for model input and output */
         R["ai_surrogate_species"] =
@@ -558,23 +685,53 @@ int main(int argc, char *argv[]) {
         const std::string ai_surrogate_input_script =
             init_list.getChemistryInit().ai_surrogate_input_script;
 
-        MSG("AI: sourcing user-provided script");
+        MSG("AI: Sourcing user-provided script");
         R.parseEvalQ(ai_surrogate_input_script);
+        if (!Rcpp::as<bool>(R.parseEval("exists(\"model_file_path\")"))) {
+          throw std::runtime_error("AI surrogate input script must contain a value for model_file_path");
+        }
 
-        MSG("AI: initialize AI model");
-        R.parseEval("model <- initiate_model()");
-        R.parseEval("gpu_info()");
+        /* AI surrogate training and inference parameters. (Can be set by declaring a 
+        variable of the same name in one of the the R input scripts)*/
+        run_params.training_data_size = init_list.getDiffusionInit().n_rows *
+                                        init_list.getDiffusionInit().n_cols; // Default value is number of cells in field
+        if (Rcpp::as<bool>(R.parseEval("exists(\"batch_size\")"))) {
+          run_params.batch_size = R["batch_size"];
+        }
+        if (Rcpp::as<bool>(R.parseEval("exists(\"training_epochs\")"))) {
+          run_params.training_epochs = R["training_epochs"];
+        }
+        if (Rcpp::as<bool>(R.parseEval("exists(\"training_data_size\")"))) {
+          run_params.training_data_size = R["training_data_size"];
+        }
+        if (Rcpp::as<bool>(R.parseEval("exists(\"use_Keras_predictions\")"))) {
+          run_params.use_Keras_predictions = R["use_Keras_predictions"];
+        }
+        if (Rcpp::as<bool>(R.parseEval("exists(\"disable_training\")"))) {
+          run_params.disable_training = R["disable_training"];
+        }
+        if (Rcpp::as<bool>(R.parseEval("exists(\"save_model_path\")"))) {
+          run_params.save_model_path = Rcpp::as<std::string>(R["save_model_path"]);
+          MSG("AI: Model will be saved as \"" + run_params.save_model_path + "\"");
+        }
+        if (Rcpp::as<bool>(R.parseEval("exists(\"use_k_means_clustering\")"))) {
+          run_params.use_k_means_clustering = R["use_k_means_clustering"];
+          MSG("K-Means clustering will be used for the AI surrogate")
+        }
+        if (Rcpp::as<bool>(R.parseEval("exists(\"train_only_invalid\")"))) {
+          run_params.train_only_invalid = R["train_only_invalid"];
+        }
+        if (!Rcpp::as<bool>(R.parseEval("exists(\"model_reactive_file_path\")"))) {
+          R.parseEval("model_reactive_file_path <- model_file_path");
+        }
+        MSG("AI: Initialize Python for the AI surrogate functions");
+        std::string python_keras_file = std::string(SRC_DIR) +
+          "/src/Chemistry/SurrogateModels/AI_Python_functions/keras_AI_surrogate.py";
+        Python_Keras_setup(python_keras_file, run_params.cuda_src_dir);
       }
 
       MSG("Init done on process with rank " + std::to_string(MY_RANK));
 
-      // MPI_Barrier(MPI_COMM_WORLD);
-
-      DiffusionModule diffusion(init_list.getDiffusionInit(),
-                                init_list.getInitialGrid());
-
-      chemistry.masterSetField(init_list.getInitialGrid());
-
       Rcpp::List profiling = RunMasterLoop(R, run_params, diffusion, chemistry);
 
       MSG("finished simulation loop");

src/poet.hpp.in

@@ -29,11 +29,13 @@
 
 #include <Rcpp.h>
 
+#define SRC_DIR "@CMAKE_SOURCE_DIR@"
+#define CUDA_SRC_DIR "@CUDA_TOOLKIT_ROOT_DIR@"
+
 static const char *poet_version = "@POET_VERSION@";
 
 // using the Raw string literal to avoid escaping the quotes
 static const inline std::string kin_r_library = R"(@R_KIN_LIB@)";
-
 static const inline std::string init_r_library = R"(@R_INIT_LIB@)";
 static const inline std::string ai_surrogate_r_library =
     R"(@R_AI_SURROGATE_LIB@)";
@@ -67,5 +69,15 @@ struct RuntimeParameters {
   static constexpr std::uint32_t INTERP_BUCKET_ENTRIES_DEFAULT = 20;
   std::uint32_t interp_bucket_entries;
 
-  bool use_ai_surrogate = false;
+  /*AI surriogate configuration*/
+  bool use_ai_surrogate = false; // Can be set with command line flag ---ai-surrogate
+  bool disable_training = false; // Can be set in the R input script
+  bool use_k_means_clustering = false; // Can be set in the R input script
+  bool use_Keras_predictions = false; // Can be set in the R input script
+  bool train_only_invalid = false; // Can be set in the R input script
+  int batch_size = 2560; // default value determined in test on the UP Turing cluster
+  int training_epochs = 20;; // Can be set in the R input script
+  int training_data_size; // Can be set in the R input script
+  std::string save_model_path = ""; // Can be set in the R input script
+  std::string cuda_src_dir = CUDA_SRC_DIR; // From CMake
 };