Comparison of semi-empirical models, symbolic regression, and machine learning approaches for prediction of tensile strength in steels

We employ data-driven models to predict the tensile strength of steel coils using information on their chemical composition and process parameters. The dataset contains extensive chemical analyses, diverse process parameters, and the characterized tensile strength as target property. We compare prediction quality and traceability of the predictions of pure machine learning models and physics-informed models. To introduce physical knowledge, we combine models from literature knowledge with symbolic regression and compare the physics-inspired models to machine learning models. In contrast to classic black-box models, symbolic regression provides mathematical equations for the estimation of the target value, facilitating straightforward interpretation. To analyze the predictions from classic black-box machine learning models, we use feature importance analysis with SHAP and contrast the obtained feature impacts with physics-based model parameters. We find that for the present use case, Artificial Neural Networks perform best, while the physics-infused models from symbolic regression allow for better interpretability.