Accelerating PINN Training via RL-based Adaptive Loss Control

Vladislav Kolzhetsov, Andrei Zakharov, Ilya Makarov

Keywords: PINNs, Reinforcement Learning, Adaptive Loss Balancing, PDE Learning, Scientific ML

TL;DR: We use reinforcement learning to dynamically control PINN loss weights, significantly accelerating convergence without sacrificing accuracy

Abstract: Physics-Informed Neural Networks have demonstrated strong performance across scientific and engineering applications by integrating governing physical laws, expressed as partial differential equations, into neural training. However, their practical use is often limited by slow and unstable training. A key factor is the balancing multiple terms within the loss function, typically with static, user-defined weighting coefficients. In this work we show that PINN training can be sufficiently accelerated by employing a reinforcement learning agent to dynamically adapt these loss weights. We formulate loss balancing as a sequential decision-making problem and demonstrate that an RL-driven policy reduces the number of training iterations required to reach a target accuracy compared to static weighting schemes. Numerical experiments on canonical PDE benchmarks show consistent convergence acceleration without loss of solution fidelity, highlighting the potential of adaptive loss control for more efficient physics-informed learning.

Journal Opt In: Yes, I want to participate in the IOP focus collection submission

Submission Number: 68