Modularity-Free Conflict-Averse Training for Generalized PINNs

The paper introduces a new training framework called Modular-Sparsity Synchronization (ModSync) for Physics-informed Neural Networks (PINNs) that addresses the challenges of training overparameterized models, which can suffer from functional modularity and hinder convergence. ModSync integrates structural optimization to penalize task-exclusive connections while enhancing cross-objective interactions, leading to improved training stability and state-of-the-art accuracy across various PDE benchmarks. This advancement is significant for practitioners as it enables more effective use of high-capacity models in solving complex physical problems without the typical training fragility associated with PINNs.