Physics-informed ANNs for dynamic, distributed and stochastic systems (SmartANN)

We rate two aspects of machine learning (ML) as extremely critical: (1) it ignores the physical process understanding accumulated in past centuries and (2) ML results are too black-box-like for providing logical explanations in building new process understanding. Our goal is that scientific understanding can assist ML, and that ML becomes useful in the quest for new process understanding and providing explanations. To achieve this goal, we design appropriate forms of artificial neural networks (ANNs) with a physics-informed internal structure and hypothesis space. Then, we confine their learning process with additional physics-based knowledge. Finally, we teach them to realize their own tendency for prediction errors and to provide this insight in the form of statistical distributions. This way, we obtain stochastic ML models that are not a black box, have better learning success from less training examples, achieve superior forecasting skills with honest uncertainty intervals. Additionally, they can be used accurately for data assimilation from systems that are understood only in parts. We develop and validate our ideas on a prototypical energy storage device. In this application, we use ML in order to track the state of charge, temperature and pressure, and to predict the state of health and remaining lifetime of the device. The German Aerospace Center (DLR) will provide experimental data from such a prototype, and partners from Project Network 1 will provide existing simulation tools.