PECANNs - Physics and Equality Constrained Artificial Neural Networks

Explore a novel approach to physics-informed neural networks (PINNs) in this 58-minute talk by Shamsulhaq Basir at the Alan Turing Institute. Delve into the limitations of traditional PINNs, particularly the challenge of balancing hyperparameters for residual forms of partial differential equations and boundary conditions. Discover how these hyperparameters significantly impact prediction accuracy and why their optimal values don't translate across different problems or training settings. Learn about the proposed physics-informed and equality-constrained framework designed to address these limitations for both forward and inverse problems. Understand how this new approach is noise-aware and capable of multi-fidelity data fusion. Examine the framework's application to various multi-dimensional PDE problems, showcasing its efficacy in achieving substantial improvements in accuracy compared to state-of-the-art PINNs. Gain insights into this innovative method that promises to enhance the reliability and versatility of physics-informed neural networks in scientific applications.