Figure 1
The diagram illustrates a genetic algorithm optimized loss balancing framework for physics-informed neural networks. The framework includes a genetic algorithm meta-optimization process with steps such as population initialization, fitness evaluation, selection, crossover, and mutation. The genetic algorithm adjusts loss weights for data loss, physics residual loss, and boundary/initial condition loss. These optimized loss weights are fed back into the physics-informed neural network training loop, which receives spatial-temporal inputs and predicts the physical field.Schematic of the proposed genetic algorithm (GA)–optimized loss balancing framework for physics-informed neural networks (PINNs). The PINN receives spatial–temporal inputs and predicts the physical field . Training is guided by three loss components: data loss, physics residual loss derived from the governing equations, and boundary/initial condition loss. A genetic algorithm operates at a meta-optimization level to automatically adjust the corresponding loss weights through population initialization, fitness evaluation, selection, crossover, and mutation. The optimized loss weights are fed back into the PINN training loop, enabling systematic and physics-aware loss balancing without manual tuning