Physics-Based vs. Data-Driven Methods to Accelerate Battery Test Cycles

Dr. Richard Ahlfeld, Monolith AI

Abstract:

Dr. Richard Ahlfeld, CEO of Monolith, will cover the topic of Machine Learning to maximise the value of engineering tests. With a primary focus on a battery test case study, Dr. Ahlfeld will demonstrate the differences between physics-based and data-driven models, where physics-based approaches fall short and how Machine Learning can complement those. Validating complex, nonlinear systems is hard. And testing every possible scenario on a test bench or in a simulation is not feasible. Over-testing confirms what's already known, while under-testing risks failing certification or missing issues. Physics based models used in engineering tests rely on complex mathematical equations to simulate and predict system behaviour. While these models have been valuable, they often struggle to capture the full complexity and intricacies of real-world systems. Dr. Ahlfeld will discuss the limitations of physics-based models in dealing with uncertainties, non-linearities, and the intractability of the physics that is being investigated. Optimising multiple design parameters through time-consuming experiments poses challenges in various scientific and engineering fields. For instance, in the battery test study, maximising battery lifetime requires extensive experimentation that can take months to years, with each experiment lasting several hours. To address this, Monolith developed an early-prediction model (Next Test Recommender). The Next Test Recommender provides engineers with active recommendations for the exact best test conditions to choose from for the next batch of tests and ranks the most impactful new tests to carry out, based on an analysis of previously collected data. This approach significantly reduces the time and number of experiments required, substantially reducing traditional exhaustive search methods that would take over 500 days to just 16 days (equivalent to 384 hours).