Learning strategies for physics-informed neural networks
Jian Cheng Wong · 2025
Physics-informed neural networks (PINNs) are at the forefront of scientific machine learning, making possible the creation of machine intelligence that is cognizant of physical laws and able to accurately simulate them. This thesis studies the challenges of physics-informed learning and explores new strategies to improve it. The discussion highlights recent findings that reveal the difficulties in model training when switching from standard data-driven losses to physics-informed learning objectives. We focus on practical forward and inverse PINN models that incorporate ideas from scientific computing to improve learning efficiency and model accuracy. Additionally, we explore nature-inspired learning algorithms that can find globally optimum solutions for PINNs that adhere closely to physics laws. By combining ideas from evolutionary computation with transfer learning and meta-learning, effective gradient-free learning algorithms can be developed to improve physics-informed models across a distribution of tasks.