Incorporating Domain Knowledge in Statistical Machine Learning to Efficiently Navigate Parameter Space and Test Durability for Energy Storage

<b>Given the urgency of addressing global warming and the need for sustainable energy solutions, the importance of efficient energy storage devices cannot be overstated. Of particular interest are the long-duration storage systems that are paramount for the integration of these sustainable sources, like renewables, in the power grid. Long-duration storage systems are designed to capture surplus energy during periods of high production and low demand, subsequently releasing it during periods of low production and high demand. </b><br/><br/><b>Developing such systems requires efficiently exploring their parameter spaces and assessing durability. Traditionally, this exploration process has been haphazard and inefficient, resulting in time and resource wastage, and an inadequate assessment of durability and failure probabilities. Machine learning and artificial intelligence are revolutionizing most aspects of science and engineering, including parameter space exploration, but adopting those advancements has been slow in the energy-storage community. The reasons are the expense of battery testing and the associated sparsity of datasets, the need to extrapolate instead of interpolate, the inherent stochasticity of the problem, and the deep connection between failure mechanisms and physical and chemical processes. All of this causes a purely data-driven approach to be suboptimal. In this study, we present an innovative method that harnesses statistical and machine learning techniques, specifically Gaussian Process regression and the expected information gain from future experiments, to streamline the exploration of energy storage systems parameter space. By incorporating the expertise of domain specialists to tailor our prior mean, kernel function, and noise model, our approach minimizes the need for extensive experimentation while accurately quantifying the failure probability distribution. Our results show that this method can, in fact, efficiently explore the parameter space and approximate failure distributions early in the testing process. In short, this work holds promise for expediting the development and optimization of energy storage, facilitating renewable energy integration, and contributing to a more sustainable future.</b>