Automating Physical Knowledge Integration in Machine Learning

In a world where training data is often limited or noisy, how can we improve the reliability of machine learning models? And how can prior knowledge be integrated into these models to be closer to reality? This paper introduces Ontology-based Physics-Informed Machine Learning (OPIML), an innovative approach that formalizes and integrates physical knowledge into machine learning models using ontology. An illustration of the approach is provided with the loss function of a neural network. By using ontologies to automate the transformation of physical laws into mathematical equations, OPIML paves the way for more accurate and robust models. Experimental results demonstrate the successful application of abstract rules in various material design scenarios, offering a pathway for more robust and precise models. Later on, this approach could also be extended to encompass a variety of rule types, including legal and ethical constraints.