کنفرانس بین المللی فناوری اطلاعات و دانش

صفحه اصلی / شانزدهمین کنفرانس بین المللی فناوری اطلاعات و دانش

TDO-SA-PINN: A Co-Evolutionary Framework for Physics-Informed Neural Networks

نویسندگان :

SeyedMohammadReza AhmadEnjavi¹ Masoud Shafiee²

1- دانشگاه صنعتی امیرکبیر 2- دانشگاه صنعتی امیرکبیر

کلمات کلیدی :

Physics-Informed Neural Networks،Tasmanian Devil Optimizer،Optimization for Deep Learning

چکیده :

Physics-Informed Neural Networks (PINNs) have emerged as a promising paradigm for solving forward and inverse partial differential equations (PDEs), yet their performance often deteriorates in stiff, multi-scale, or high-frequency regimes due to spectral bias, loss imbalance, and local optimization pathologies. While Self-Adaptive PINNs (SA-PINNs) mitigate error concen tration by dynamically adjusting residual weights, their correc tive power remains constrained by gradient-based optimizers that stagnate in rugged landscapes. To address this gap, we introduce a co-evolutionary framework that integrates SA-PINNs with the Tasmanian Devil Optimizer (TDO), a recent population based metaheuristic. In the proposed TDO-SA-PINN, adaptive weights reshape the loss landscape while a diverse swarm of candidate networks performs global, gradient-free exploration. This dual mechanism simultaneously targets spectral bias and optimizer-induced stagnation, and naturally yields an ensemble that encodes predictive uncertainty. Extensive experiments on canonical PDE benchmarks demonstrate that TDO-SA-PINNs achieve lower error and more reliable convergence compared to standard PINNs trained with ADAM/LBFGS, adaptive PINN variants, and deep ensembles. The results highlight the potential of co-evolutionary population search as a scalable and effective complement to adaptive physics-informed learning frameworks.