BatteryInformatics

Source code and trained models for the paper "Comparative Analysis of Structure-Property Machine Learning models for Predicting Electrolyte Thermodynamic Windows".

General overview of the modeling framework

Table of Contents

Table of Contents

1. ➤ Set-up environment
2. ➤ Training Machine Learning Models
   • Traditional Model
   • Graph Neural Network
   • Transformer
3. ➤ Status
4. ➤ How to cite
5. ➤ License
6. ➤ References

Set-up environment

Clone this repository and then use setup.sh to setup a virtual environment binfo with the required dependencies in requirements.txt.

chmod +x setup.sh
git clone https://github.com/EnthusiasticTeslim/BatteryInformatics.git
cd BatteryInformatics
sh setup.sh
source binfo/bin/activate

Training Models

Important

All scripts for training models are available in Docker mode in folder docker.

Traditional Model

python src/descriptor/trainer.py -h
usage: trainer.py [-h] [--parent_directory PARENT_DIRECTORY] [--data_directory DATA_DIRECTORY] [--result_directory RESULT_DIRECTORY] [--src SRC] [--train_data TRAIN_DATA] [--test_data TEST_DATA] [--scale] [--hyperparameter HYPERPARAMETER] [--iterations ITERATIONS] [--cv CV] [--model MODEL] [--seed SEED]

options:
  --parent_directory PARENT_DIRECTORY
                        Path to main directory
  --data_directory DATA_DIRECTORY
                        where the data is stored in parent directory
  --result_directory RESULT_DIRECTORY
                        Path to result directory
  --src SRC             function source directory
  --train_data TRAIN_DATA
                        Path to train data
  --test_data TEST_DATA
                        Path to test data
  --scale               Scale data
  --hyperparameter HYPERPARAMETER
                        Hyperparameter space
  --iterations ITERATIONS
                        Number of iterations for hyperparameter optimization
  --cv CV               Number of cross-validation folds
  --model MODEL         Model to train
  --seed SEED           Random seed
  --morgan_fingerprint  Use Morgan Fingerprint (MFF) instead of RDKit descriptors
  --nbits NBITS         Number of bits for MFF
  --radius RADIUS       Radius for MFF

The model and its predictions will be saved in results/<MODEL>. For example, to train a SVR model using RDKIT descriptor, you can use:

python -m src/descriptor/trainer.py --parent_directory YOUR_MAIN_FOLDER --result_directory results --data_directory data --train_data "train_data_cleaned.csv" --test_data "test_data_cleaned.csv" --scale --model SVR --seed 42 --iterations 100 --hyperparameter "hp_descriptor.yaml" --cv 5

To train the whole model (SVR, RandomForest, AdaBoostRegressor, GradientBoostingRegressor),

chmod a+x regenerate/descriptor.sh
./descriptor.sh

Graph Neural Network

python src/graph/trainer.py -h
trainer.py [-h] [--parent_directory PARENT_DIRECTORY] [--result_directory RESULT_DIRECTORY] [--data_directory DATA_DIRECTORY] [--train_data TRAIN_DATA] [--test_data TEST_DATA] [--add_features]
                  [--skip_cv] [--epochs EPOCHS] [--start-epoch START_EPOCH] [--batch_size BATCH_SIZE] [--lr LR] [--gpu GPU] [--cv CV] [--dim_input DIM_INPUT] [--unit_per_layer UNIT_PER_LAYER] [--seed SEED]
                  [--num_feat NUM_FEAT] [--train]

options:
  --parent_directory PARENT_DIRECTORY
                        Path to main directory
  --result_directory RESULT_DIRECTORY
                        Path to result directory
  --data_directory DATA_DIRECTORY
                        where the data is stored in parent directory
  --train_data TRAIN_DATA
                        name of train data
  --test_data TEST_DATA
                        name of test data
  --add_features        if add features
  --skip_cv             if skip cross validation
  --epochs EPOCHS       number of total epochs to run
  --start-epoch START_EPOCH
                        manual epoch number (useful on restarts)
  --batch_size BATCH_SIZE
                        mini-batch size (default: 256)
  --lr LR               initial learning rate
  --gpu GPU             GPU ID to use.
  --cv CV               k-fold cross validation
  --dim_input DIM_INPUT
                        dimension of input
  --unit_per_layer UNIT_PER_LAYER
                        unit per layer
  --seed SEED           seed number
  --num_feat NUM_FEAT   number of additional features
  --train               if train
  --print_result        if print result

To train the GNN model,

chmod a+x regenerate/graph.sh
./graph.sh

and its checkpoints and predictions will be saved in results/GNN.

For example, to train a GNN model you can use:

python -m src/graph/trainer.py --parent_directory YOUR_MAIN_FOLDER --result_directory results --data_directory data --train_data "train_data_cleaned.csv" --test_data "test_data_cleaned.csv" --seed 42 --iterations 100 --train --cv 5

to test and an already train model, you can use:

python -m src/graph/trainer.py --parent_directory YOUR_MAIN_FOLDER --result_directory results --data_directory data --train_data "train_data_cleaned.csv" --test_data "test_data_cleaned.csv" --seed 42 --iterations 100 --cv 5

Transformer

under construction

Status

• Complete data cleaning
• Scripts for QSPR with RDKIT descriptors.
• Scripts for QSPR with Graph.
• Train ML with RDKIT and Graph.
• [] Set up ML model with transformer.
• [] Evaluate performances.
• [] Deploy models as a GUI.

How to cite

@article{doi,
  author = {Teslim Olayiwola, Jose Romagnoli},
  title = {Comparative Analysis of Structure-Property Machine Learning models for Predicting Electrolyte Thermodynamic Windows},
  journal = {n/a},
  year = {n/a},
  volume = {n/a},
  number = {n/a},
  doi = {https://doi.org/},
  preprint = {Manuscript in Preparation}
}

License

BatteryInformatics is under MIT license. For use of specific models, please refer to the model licenses found in the original packages.

References

• Qin et al (2021) Predicting Critical Micelle Concentrations for Surfactants Using Graph Convolutional Neural Networks
• Miriam et al (2024) Surfactant-Specific AI-Driven Molecular Design: Integrating Generative Models, Predictive Modeling, and Reinforcement Learning for Tailored Surfactant Synthesis