AI for Non-aqueous Electrolyte Design

This software package implements the SCAN (Shaping Conductivity Atlas for Non-aqueous electrolytes) that takes Li-salts, solvents, and conditions to predict the ionic conductivity.

The package provides three major functions:

• Calculate the descriptors based on Li-salts, solvents, and conditions.
• Train a SCAN model with the entire data.
• Predict conductivity based on SCAN model or symbolic regression model.

Prerequisites

• torch==2.2.1
• rdkit==2024.3.6
• scikit-learn==1.5.1
• pysr==1.5.2

The easiest way of installing the prerequisites is via conda. After installing conda, run the following command to create a new environment named scan and install all prerequisites:

conda upgrade conda
conda create -n scan python=3.12 scikit-learn pytorch rdkit pysr

This creates a conda environment for running SCAN. Before using SCAN, activate the environment by:

source activate scan

Alternatively, environment.yaml provides the dependencies for creating running environment. Then, in directory model, you can test if all the prerequisites are installed properly by running:

python train.py

After you finished using SCAN, exit the environment by:

source deactivate

Models

We provide the model files in model and sampling directories.

• sampling: the original multi-feature network (MFNet) is provided, which is one of the baseline model. Additionally, three strategies (over-sampling, under-sampling, hybrid-sampling) with various hyper-parameters were adopted for training baseline models. To run these models:

    python over_sampling_train.py
    python under_sampling_train.py
    python over_under_sampling_train.py
  

• model: MFNet with dynamic routing strategy was implemented for predicting the conductivity, to run the model:

    python train.py
  

• trained_pth_model: trained model weights of MFNet models and symbolic regression models

    MFNet_fold_i_model.pth # well-trained NFNet models using five-fold cross validations
    checkpoint_unary.pkl # symbolic regression model, with unary and binary operations
    checkpoint_binary.pkl # symbolic regression model, with only binary operations
  

Scripts

We provide the practical scripts in examples to predict the ionic conductivity based on well-trained models or symbolic regression models, and in utils directory for calculating the ionic conductivity, constructing simulation box, and calculating molecular properties.

• Ionic conductivity prediction

You can predict ionic conductivity based on MFNet models, by running:

    python predict.py

If you want to predict ionic conductivity based on symbolic regression models, run:

    python symbolic_regression.py

• Molecular property calculation

If you want to re-calcualte or re-design the molecular properties, just provide the SMILES for a given molecule, and run:

    python molecular_property.py

• Simulation box construction

This tool is useful to calculate the number of molecules for Li-salts and two solvents in the simulation box, according to parameters: box_size, density, salt_concentration, salt_molar_mass, solvent1_molar_mass, solvent2_molar_mass, solvent_mass_ratio, by running:

    python box_construction.py

• Conductivity calculation

After obtaining the diffusion coefficent from the MD simultions, you can use this tool to calculate the ionic conductivity based on Arrhenius equation:

    python conductivity_calculation.py

Data

To reproduce our paper, you can download the corresponding datasets in data directory.

• calisol: lists the compiled data, including k values, temperature, concentration/unti, salt, solvent. The temperature was scaled by a factor of 100.
• salt_feature.npy: feature matrix of salt molecules based on the designed descriptor.
• solvent_feature.npy: feature matrix of solvent molecules based on the designed descriptor.
• condition_feature.npy: feature matrix of conditions.
• conductivity_target.txt: collected k values.

Author contributions

This software was primarily written by Dr. Zhilong Wang who is advised by Prof. Fengqi You.

Web app

A online platform was estabilished that allows researchers to query our non-aqueous electrolyte database and predict conductivity properties using our deep learning models. It accelerates the discovery of non-aqueous electrolyte for battery and energy storage applications.

    https://peese-scan.streamlit.app/

SCAN.mp4

How to cite

Please cite the following work if you want to use SCAN:

Zhilong Wang, Fengqi You*. Nature Computational Science Accepted in principle (2025).