The basic idea of machine learning is that an artificial system learns from patterns and relations in data. This approach has become very important in recent years, especially in connection with developments in deep learning. However, large volumes of data are needed to train an ML model. These are often not available in the production environment because the collection and annotation of data is invariably extremely expensive and time-consuming.
One way of solving this problem is to couple machine learning with existing knowledge. Especially in production, prior knowledge about the problem to be solved already exists. Such knowledge can be, for example, analytical models for describing systems and processes, simulations of individual processes, or even entire factories, rule-based systems, or information acquired by employees through experience.
The research area “physics-informed machine learning” focuses on ways to combine prior knowledge in the form of physical laws and equations with machine learning. By using this prior knowledge, the search space can be narrowed down considerably, and training ML models becomes much more efficient. The result are more accurate and robust predictions, often with much less data being required.