Compositional Meta-Learning for Mitigating Task Heterogeneity in Physics-Informed Neural Networks
The paper introduces LAM-PINN, a compositional meta-learning framework designed to address task heterogeneity in physics-informed neural networks (PINNs) when solving parameterized partial differential equations (PDEs). By clustering tasks based on learning affinity and using specialized subnetworks with a shared meta-network, LAM-PINN enables selective module reuse, reducing negative transfer and improving generalization. It achieves a 19.7-fold reduction in mean squared error on unseen tasks with only 10% of the training iterations required by standard PINNs.
- ▪LAM-PINN uses task-specific learning dynamics and learning-affinity metrics to cluster similar PDE tasks.
- ▪The model combines cluster-specialized subnetworks with a shared meta-network and learns routing weights for modular reuse.
- ▪It achieves a 19.7-fold lower MSE on unseen tasks using only 10% of the training iterations compared to conventional PINNs.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2604.26999 (cs) [Submitted on 29 Apr 2026] Title:Compositional Meta-Learning for Mitigating Task Heterogeneity in Physics-Informed Neural Networks Authors:Beomchul Park, Minsu Koh, Heejo Kong, Seong-Whan Lee View a PDF of the paper titled Compositional Meta-Learning for Mitigating Task Heterogeneity in Physics-Informed Neural Networks, by Beomchul Park and 2 other authors View PDF HTML (experimental) Abstract:Physics-informed neural networks (PINNs) approximate solutions of partial differential equations (PDEs) by embedding physical laws into the loss function. In parameterized PDE families, variations in coefficients or boundary/initial conditions define distinct tasks.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv.org.
Discussion
0 commentsMore from arXiv.org
