Modeling and simulation of physical systems provide insights into their design, troubleshooting, and predictive maintenance. For decades, the primary tools used in physical systems modeling have been based on numerical and analytical methods, which are prohibitively expensive for highly complex systems. Machine learning became a promising alternative in such cases, leveraging the increasing availability of observational data. Nevertheless, in many scenarios, the data remain sparse, noisy, or not representative of the whole system and lead to unreliable predictions and overall inapplicability of these methods.
Instead of choosing one or the other, we aim to integrate physical and domain knowledge into machine learning algorithms, minimizing the risks of unpredictable outcomes and resulting in models that are efficient, robust, and reliable. The methods are applied to industrial settings in the context of predictive maintenance, such as increasing the lifetime of wind turbines, or detecting operational irregularities in maritime maneuvering machinery.