revisit-severson-et-al

NOTE: Please contact Prof. Richard Braatz, braatz@mit.edu, for access to the code repository associated with the Severson et al. publication in Nature Energy (available with an academic license). This repository is not directly related to the Nature Energy paper.

This repository contains code for our work entitled "Statistical learning for accurate and interpretable battery lifetime prediction", a follow-up paper to Severson et al. A permanent archive of this work on Zenodo is available here:

Our key scripts and functions are summarized here:

• generate_voltage_arrays.m: MATLAB script that generates capacity arrays from the battery dataset and exports them to csvs (stored in /data).
• revisit-severson-et-al.ipynb: Python notebook containing most of the analysis and figure generation.
• image_annotated_heatmap.py: Helper function from matplotlib (see docstring for source).
• lifetime_charging_time/lifetime_charging_time.ipynb: Contains a mini analysis on cycle life vs. charging time.

In addition, we used three scripts for neural net model training in the nn_models directory. Calls to these scripts look like this: python mlp_base.py --n_starts 10 --lr 0.001 --rw 0.0001 --hd 10.

• mlp_base.py: Trains MLP models
• cnn_base.py: Trains CNN models without baseline subtraction
• cnn_base_nosubtract.py: Trains CNN models with baseline subtraction

These scripts are used by another script that is written in a proprietary language used to interact with a cluster at IBM; this additional script is not included.

Finally, we include three notebooks that analyze the results of MLP & CNN training in the nn_models directory. These notebooks will not run completely without all .pkl or .pt files present.

• Analyze_Results_MLP_CV.ipynb: Analyzes the results from the MLP cross-validation study and selects the best hyperparameters. Also produces intermediate outputs including mlp_predictions.json, mlp_predictions_cycavg.json, and mlp_shap_results.csv.
• MLP_FinalModel.ipynb: Runs the final model based on the results of CV and generates the MLP interpretability figure using shap. -Analyze_Results_CNN.ipynb: Analyzes the results from the CNN study (not a formal CV study) and selects the best hyperparameters. The best performing model (CNN_n20000_rw0.001_lr0.0001_do0.0.pt) is included for reference.