Neural networks at the service of batteries manufacturing: a strategy behind GIGABAT’s innovations

CIDETEC Energy Storage (CID) attended ModVal2024, the 20th Symposium on Modeling and Validation of Electrochemical Energy Technologies, held on 13-14 March 2024 in Baden, Switzerland.

This symposium is the 20th in the series of annual events initiated in 2004 by the Swiss Federal Office of Energy, focused on advances in modeling and experimental validation of all types of fuel cells, electrolyzers and rechargeable batteries, including flow batteries.

GIGABAT’s coordinators, CIDETEC Energy Storage, gave a talk on “Physics Informed Neural Network for solving Single Particle Model without using labeled data”, reasearched by Francisco Javier Méndez-Corbacho, B. Larrarte-Lizarralde, R. Parra, J. Larrain, D. del Olmo and Dr. Elixabete Ayerbe.

This topic is strictly related to the Digitalization of cell production and Data-driven process optimization, which are two crucial pillars of GIGABAT’s work, where CID aims at developing predictive analytics AI algorithms to promote the detection of batches (coated coils) with risk for reduced quality, and also at identifying and process parameters which can potentially lead to low scrap-rate and high-quality along the cell manufacturing process.

GIGABAT’s partners explained how they solved the Single Particle Model (SPM) in several ranges applying a novel machine learning technique, referred as Physics Informed Neural Networks (PINN). This constructed machine learning model can provide nearly instant and accurate solutions to the problem once it is trained by incorporating the electrochemical equations from SPM into the neural network architecture. Moreover, in contrast to classical data-driven machine learning approaches, there is no need to use experimental or previous simulation data in the training process, and the results show no performance variation while decreasing computation time 10 times compared to FEM based model.