Compositional Meta-Learning for Mitigating Task Heterogeneity in Physics-Informed Neural Networks

A team from Korea University has introduced LAM-PINN (Learning-Affinity Adaptive Modular Physics-Informed Neural Network), a compositional meta-learning framework designed to overcome a critical limitation of physics-informed neural networks (PINNs): high retraining costs when faced with varying PDE coefficients, boundary conditions, or initial conditions. Traditional PINNs embed physical laws directly into the loss function, but each new parameter variation essentially becomes a new task—requiring a separate model or risking negative transfer from a single global initialization. LAM-PINN instead clusters tasks by using brief transfer sessions to compute learning-affinity metrics, then decomposes the model into cluster-specialized subnetworks plus a shared meta network. Routing weights selectively activate the right modules for each new configuration, avoiding the pitfalls of a single shared initialization.

Tested across three PDE benchmarks, LAM-PINN delivered a 19.7-fold average reduction in mean squared error (MSE) on unseen tasks while using only 10% of the training iterations required by conventional PINNs. This efficiency makes it particularly valuable for engineering applications where computational budgets are tight—such as structural analysis, fluid dynamics, or materials design—where rapid generalization across a parameter design space is essential. The paper has been accepted by Pattern Recognition, underscoring its significance for the AI and computational science communities.