MI-PINN: Meta-Learning for Inverse ODE Problems
Researchers propose a meta-inverse physics-informed neural network (MI-PINN) to solve inverse problems in high-dimensional ordinary differential equations (ODEs). The method reformulates inverse modeling as a two-stage meta-learning problem, first learning a physics-aware representation across multiple tasks, then performing inverse modeling by optimizing task-specific parameters. This approach addresses optimization difficulties and poor generalization in existing PINNs, which rely on joint optimization. The work targets scientific machine learning applications where underlying physics is partially characterized and observations are sparse.
Key facts
- MI-PINN addresses inverse problems in high-dimensional ODEs.
- It uses a two-stage meta-learning approach.
- First stage: learns physics-aware representation across tasks.
- Second stage: optimizes task-specific parameters for inverse modeling.
- Overcomes optimization difficulties and poor generalization of existing PINNs.
- Targets scenarios with partial physics knowledge and sparse observations.
- Paper published on arXiv with ID 2605.03511.
Entities
Institutions
- arXiv