PINNProv: Provenance for Physics-Informed Neural Networks

Machine Learning is being used increasingly in different application areas. Physics-Informed Neural Networks (PINN) stand out, adapting neural networks to predict solutions to Physics phenomena. Incorporating Physics knowledge into the loss function of a neural network, PINNs revolutionize the solutions of partial differential equations. Considering the lack of support for analytics and reproducibility of the trained models, in this paper we propose the capture and use of provenance data, aimed at the analysis of PINN models. We conducted experiments using TensorFlow and DeepXDE, in a high-performance computing environment. Our experiments show the contributions of these provenance queries in different PINN applications.