Résumé
The performance of electrochemical cells for energy storage and conversion, such as batteries and fuel cells, can be improved by optimizing their manufacturing processes. This can be very time consuming and costly through the conventional trial-and-error approaches. Machine Learning (ML) models can help to accelerate finding solutions to these types of problems. In academic research laboratories, manufacturing dataset sizes can be small, while ML models typically require large amounts of data. In this work, we propose a simple Transfer Learning (TL) approach where a Neural Network (NN) is trained in a vast dataset. Then, this NN is transferred to smaller datasets by freezing its weights and adding an extra trainable layer to improve the performance of this new TL-based NN. This novel approach is tested with pre-existing manufacturing experimental and stochastically generated datasets that were not acquired with the purpose of training ML models.