PEGNet: A Physics-Embedded Graph Network for Long-Term Stable Multiphysics Simulation

The introduction of PEGNet marks a significant advancement in the field of multiphysics simulation, addressing the limitations of traditional numerical solvers that are often computationally intensive. By embedding essential PDE dynamics into its architecture, PEGNet utilizes a novel approach of PDE-guided message passing, which enhances the stability and physical consistency of simulations. This method is particularly beneficial in complex geometries where conventional data-driven approaches struggle with error accumulation. Evaluated on benchmarks for respiratory airflow and drug delivery, PEGNet demonstrated substantial improvements in long-term prediction accuracy, showcasing its potential to transform how physical phenomena are modeled and understood. The integration of physical regularization into the loss function further reinforces its adherence to governing physics, making it a promising tool for scientific and engineering advancements.