Abstract
BridgeNet is a novel hybrid framework that integrates convolutional neuralnetworks with physics-informed neural networks to efficiently solve non-linear,high-dimensional Fokker-Planck equations (FPEs). Traditional PINNs, whichtypically rely on fully connected architectures, often struggle to capturecomplex spatial hierarchies and enforce intricate boundary conditions. Incontrast, BridgeNet leverages adaptive CNN layers for effective local featureextraction and incorporates a dynamically weighted loss function thatrigorously enforces physical constraints. Extensive numerical experimentsacross various test cases demonstrate that BridgeNet not only achievessignificantly lower error metrics and faster convergence compared toconventional PINN approaches but also maintains robust stability inhigh-dimensional settings. This work represents a substantial advancement incomputational physics, offering a scalable and accurate solution methodologywith promising applications in fields ranging from financial mathematics tocomplex system dynamics.