This Two-Stage AI Trick Just Made PINNs 40% More Accurate – Here's the Secret

Physics-Informed Neural Networks (PINNs) solve partial differential equations (PDEs) by embedding physical laws into training, but their performance is notoriously sensitive to hyperparameters due to a highly non-convex and multi-term loss function. Traditional gradient-based or local search methods often get trapped in suboptimal regions. A new paper from HSE University introduces a two-stage evolutionary strategy that first uses population-based algorithms to rapidly explore the hyperparameter space with truncated training epochs, then refines the best candidates with full gradient-based training. This outer-loop approach treats hyperparameter selection as a black-box optimization problem, leveraging evolutionary algorithms' ability to handle mixed, non-differentiable search spaces.

Evaluated on three classic PDE benchmarks—Advection, Klein-Gordon, and Helmholtz equations—the method consistently outperformed standard training, achieving significantly lower mean error within the same computational budget. The two-stage design balances exploration (wide search) and exploitation (deep refinement), making PINNs more robust and practical for real-world scientific computing tasks. Accepted at the ICLR 2026 Workshop on AI and PDEs, this work provides a scalable framework for automating PINN hyperparameter tuning, potentially accelerating adoptiers in engineering and physics simulations where accurate PDE solvers are critical.