Continual Learning for Accelerating Flow Battery Redox-Active Material Development

Redox flow batteries hold great promise for large-scale energy storage applications due to their scalable capacity and low fire risks. The machine learning-based regression models have demonstrated their capability to provide high throughput performance predictions as a function of the battery cell design, operating conditions, and relevant material properties. However, due to the large design space of the molecular especially for the emerging organic redox-active materials, new datasets to be used for algorithm training can be available only incrementally. Moreover, it is computationally inefficient to simply add them to the existing material datasets and retrain the machine learning algorithms. To remedy these difficulties and help to accelerate the redox-active material design, we have developed continual learning methods for flow battery predictions that can train on incremental datasets, without a replay of previous datasets. Our multifidelity continual learning method takes advantage of correlations between the datasets to improve predictions without catastrophic forgetting that can occur for the earliest datasets. This approach can be combined with existing continual learning techniques, such as elastic weight consolidation, and other neural network training methods, such as physics-informed neural networks, to improve the flow battery performance predictions.