LECA - The Liquid Electrolyte Composition Analysis Package

The Liquid Electrolyte Composition Analysis (LECA) package creates a simplified Jupyter-Notebook [1] based workflow for applying Scikit-Learn [2] machine learning regression models to predict liquid electrolyte behavior based on composition.

Requirements

Python 3.7+, Jupyter Notebook 6.4.11+

With the following python libraries:

• Scikit-learn 1.3.1+
• NumPy 1.22.3+
• Matplotlib 3.5.1+
• Pandas 1.4.2+
• SciPy 1.8.1+
• Uncertainties 3.1.7+
• MAPIE 0.5.6
• HDBSCAN 0.8.28+
• Seaborn 0.11.2+
• GPyOpt 1.2.6+

Installation

The LECA package can be installed directly from source with:

pip install git+https://github.com/Harrison-Teeg/LECA.git

Envisaged LECA Work Flow

Data Import and Feature Engineering

• Import Data from JSON or CSV files
• Visualize dataset with feature overview and interactive data visualizer
• Identify and filter outlier values using HDBSCAN [3]
• Manually filter nonsense-values with user defined explicit boundaries
• Combine repeated measurements and record statistical behavior (measurement noise)
• Generate surrogate models for Arrhenius behavior or other user defined values

Initialize Regression Models and Compare Results

• Data splitting / Scaling automatically handled

• Declare regression models to implement (supports N-dimensional feature/objective space)

  • Linear Regression (LR)
  • Gaussian Process Regression (GPR) (supports Isotropic/Anisotropic RBF, Matern, RQ, Custom kernel)
  • Neural Network (NN)
  • Random Forest (RF)

• Hyperparameter Optimization for NN and RF with GPyOpt [4]
• Customized Polynomial selection for LR [5]
• Cross-validated scoring of models and visualization to provide simple overview of comparative model performance
• Ensemble based uncertainty estimation for LR / NN / RF models using MAPIE [6]
• Validate performance of models on unseen validation data

Analyze Objective Function for Compositions

• Interactive widgets to visualize objective function and model uncertainty for various compositions
• Return optimal composition to maximize/minimize objective function optimization
• Ranked Batch Mode Active Learning module based on RBMAL approach of Cordoso et al. [7]

Areas of Further Development

Multi-Objective Optimization: Identifying Pareto-fronts for multiple-objectives for electrolyte composition (e.g. electrochemical stability, conductivity, etc.)

References

[1]	[9] Brian E. Granger and Fernando Pérez. “Jupyter: Thinking and Storytelling With Code and Data”. In: Computing in Science & Engineering 23.2 (2021), pp. 7–14. doi: 10.1109/MCSE.2021.3059263.

[2]	Pedregosa et al. “Scikit-learn: Machine Learning in Python”. In: Journal of Machine Learning Research 12 (2011), pp. 2825–2830.

[3]	Leland McInnes, John Healy, and Steve Astels. “hdbscan: Hierarchical density based clustering”. In: The Journal of Open Source Software 2.11 (2017), p. 205.

[4]	The GPyOpt authors. GPyOpt: A Bayesian Optimization framework in python. 2016. url: http://github.com/SheffieldML/GPyOpt.

[5]	Anand Narayanan Krishnamoorthy et al. “Data-Driven Analysis of High-Throughput Experiments on Liquid Battery Electrolyte Formulations: Unraveling the Impact of Composition on Conductivity**”. In: Chemistry–Methods 2.9 (2022), e202200008. doi: https://doi.org/10.1002/cmtd.202200008.

[6]	MAPIE - Model Agnostic Prediction Interval Estimator. Version: 0.4.1. url: https://mapie.readthedocs.io/en/latest/index.html (visited on 08/24/2022).

[7]	Thiago N.C. Cardoso et al. "Ranked batch-mode active learning". In: Information Sciences 379 (2017) pp. 313-337. doi: https://doi.org/10.1016/j.ins.2016.10.037

Acknowledgments

This project has received funding from the European Union’s Horizon 2020 research and innovation program under grants agreement No 957189 (BIG-MAP) and No 957213 (BATTERY2030+).