PSO-PINN: Physics-Informed Neural Networks Trained with Particle Swarm Optimization

Physics-informed neural networks (PINN) have recently emerged as a promising application of deep learning in a wide range of engineering and scientific problems based on partial differential equation (PDE) models. However, evidence shows that PINN training by gradient descent displays pathologies that often prevent convergence when solving PDEs with irregular solutions. In this paper, we propose the use of a particle swarm optimization (PSO) approach to train PINNs. The resulting PSO-PINN algorithm not only mitigates the undesired behaviors of PINNs trained with standard gradient descent but also presents an ensemble approach to PINN that affords the possibility of robust predictions with quantified uncertainty. We also propose PSO-BP-CD (PSO with Back-Propagation and Coefficient Decay), a hybrid PSO variant that combines swarm optimization with gradient descent, putting more weight on the latter as training progresses and the swarm zeros in on a good local optimum. Comprehensive experimental results show that PSO-PINN with the proposed PSO-BP-CD algorithm outperforms PINN ensembles trained with other PSO variants or with pure gradient descent.