Theory-Guided Data-Science for Battery Modeling: Insights from a Comparative Study

Theory-guided data science (TGDS) uses problem-specific domain knowledge to ensure physical consistency and enhance the performance of data science models. However, despite the increasing popularity of TGDS, there is a lack of studies comparing TGDS methods on concrete data science tasks, and their deployment remains difficult. To address this gap, we present a comparative study of existing TGDS paradigms applied to time-series prediction, in the context of dynamical systems modeling. Our concrete use case is lithium-ion battery modeling. Applying TG DS paradigms in this context is not trivial, and we describe how to organize existing domain knowledge and to implement the TGDS paradigms. We uniquely introduce and evaluate essential criteria for TGDS methods (domain knowledge and intellectual effort), often missing in use-case-specific publications. Our study findings reveal that while effective TGDS methods require substantial domain knowledge, their implementation does not necessarily require significant effort.