Projects
Machine Learning on Physics-Based Model Generated Data
I’m working on a project (Repository) that uses data described in Dubarry and Beck’s 2020 Journal of Power Sources article, Big data training data for artificial intelligence-based Li-ion diagnosis and prognosis.
The data is model-generated (“synthetic”) C/25 charge curves for a graphite-LFP cell. Three different degradation modes (loss of LFP, loss of graphite, and loss of lithium inventory in the electrolyte) are simulated. The data can be found here
I’m using these data to train machine learning models for two purposes:
- Parameter estimation. ML trained on physically constrained training sets like the one above can serve as faster surrogate models to physics-based models, as described here and here.
- Classification of degradation modes
This work is an opportunity for me to learn about and implement ML algorithms. For instance, I explore the role of the nature of the training data, the size of the training data, and the preprocessing steps on performance of a Deep Neural Network for parameter regression. I also take some inspiration from Severson and Attia’s work on transforming and featurizing voltage profiles as inputs for ML models, and using interpretable ML.
BatteryDEV: Lifetime prediction
I participated in a hackathon of sorts organized by BatteryDEV and Battery Associates, where I worked with a team to build a model to predict the remaining capacity of an electric bus battery. Post.