HyPINO: Multi-Physics Neural Operators via HyperPINNs and the Method of Manufactured Solutions

2025-09-05

AI Summary

Zero-shot generalization for multi-physics parametric partial differential equations (PDEs) remains challenging due to disparities across equation types (elliptic/hyperbolic/parabolic), geometries, and boundary conditions. Method: We propose HyPINO—the first multi-physics neural operator enabling zero-shot generalization across PDE classes, domains, and boundary conditions without task-specific fine-tuning. Its core innovations include: (1) a Swin Transformer-based hypernetwork that generates physics-informed neural networks; (2) hybrid training combining analytically derived labels from manufactured solutions with physics-constrained unsupervised losses; and (3) an error-feedback-driven iterative refinement mechanism for high-accuracy forward inference. Results: On seven benchmark tasks, HyPINO reduces zero-shot L₂ error by over two orders of magnitude on average, outperforming U-Net, Poseidon, and PINO. Moreover, it serves as a high-quality initialization that accelerates subsequent PINN fine-tuning while improving both convergence speed and final accuracy.

Technology Category

• Machine Learning: Hardware-aware ML
• Computer Vision: Low Level & Physics-based Vision
• Search and Optimization: Mixed Discrete/Continuous Search

Application Category

• Economics, Online Markets and Human Computation: Incentives in network design for Web infrastructures and ecosystems
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs

Abstract

We present HyPINO, a multi-physics neural operator designed for zero-shot generalization across a broad class of parametric PDEs without requiring task-specific fine-tuning. Our approach combines a Swin Transformer-based hypernetwork with mixed supervision: (i) labeled data from analytical solutions generated via the Method of Manufactured Solutions (MMS), and (ii) unlabeled samples optimized using physics-informed objectives. The model maps PDE parametrizations to target Physics-Informed Neural Networks (PINNs) and can handle linear elliptic, hyperbolic, and parabolic equations in two dimensions with varying source terms, geometries, and mixed Dirichlet/Neumann boundary conditions, including interior boundaries. HyPINO achieves strong zero-shot accuracy on seven benchmark problems from PINN literature, outperforming U-Nets, Poseidon, and Physics-Informed Neural Operators (PINO). Further, we introduce an iterative refinement procedure that compares the physics of the generated PINN to the requested PDE and uses the discrepancy to generate a "delta" PINN. Summing their contributions and repeating this process forms an ensemble whose combined solution progressively reduces the error on six benchmarks and achieves over 100x gain in average $L_2$ loss in the best case, while retaining forward-only inference. Additionally, we evaluate the fine-tuning behavior of PINNs initialized by HyPINO and show that they converge faster and to lower final error than both randomly initialized and Reptile-meta-learned PINNs on five benchmarks, performing on par on the remaining two. Our results highlight the potential of this scalable approach as a foundation for extending neural operators toward solving increasingly complex, nonlinear, and high-dimensional PDE problems with significantly improved accuracy and reduced computational cost.

Problem

• Solving parametric PDEs with zero-shot generalization across diverse physical scenarios
• Handling varying geometries and boundary conditions without task-specific fine-tuning
• Improving accuracy and reducing computational cost for complex nonlinear PDEs

Innovation

• Uses hypernetwork with Swin Transformer for parametric PDE mapping
• Combines manufactured solutions with physics-informed loss supervision
• Implements iterative refinement with delta PINNs for error reduction