Newton-PINet: A fast physics-informed neural network with Newton linearization for meta-learning nonlinear PDEs
Abstract
Scientific machine learning has opened new avenues for solving parameterized partial differential equations (PDEs), enabling models to learn a family of PDEs and generalize to unseen instances. In this context, data-driven operator learning methods typically require large training datasets, while physics-informed neural networks (PINNs) suffer from difficult optimization and limited generalization, especially for nonlinear PDEs. We propose Newton-PINet, a physics-informed network enhanced by Newton linearization, offering an effective meta-learning framework for nonlinear PDEs. Newton-PINet (i) employs a physics-informed multilayer network with skip connections, where the output-layer weights are solved by least squares; (ii) adopts a two-stage learning strategy that first leverages gradient-based training to learn robust representations from the available training tasks, and then performs gradient-free fine-tuning on the output layer for fast task-specific generalization; and (iii) incorporates a Newton linearization method to speed up the least-squares iteration for nonlinear PDE problems. On a challenging nonlinear reaction-diffusion benchmark, Newton-PINet achieves up to three orders of magnitude lower relative error than recent neural solvers, while using 16x fewer training tasks and over an order of magnitude less training time (about 5 minutes versus several hours). This work advances the meta-learning of PINNs toward data-efficient, fast, and generalizable physics solvers. The code and datasets are publicly available at: https://github.com/fycAIwj/Newton-PINet.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
Newton-PINet introduces a novel meta-learning framework that combines Newton linearization with Tikhonov-regularized least-squares to achieve fast, data-efficient, and accurate solutions for nonlinear PDEs. By decoupling representation learning from task-specific adaptation and leveraging the superior convergence properties of Newton's method over Picard iteration, the paper demonstrates substantial improvements in both accuracy and computational efficiency compared to existing PINN and neural operator approaches, establishing a new benchmark for physics-informed meta-learning.
Methodology
The paper proposes Newton-PINet, a meta-learning framework for solving nonlinear PDEs. The core innovation is the integration of Newton linearization into a physics-informed neural network (PINN) architecture that utilizes skip connections and Tikhonov-regularized least-squares for the output layer. By linearizing nonlinear terms via Taylor expansion (Newton's method) rather than simple lagging (Picard iteration), the method achieves near-quadratic convergence for the inner-loop adaptation. The architecture separates the heavy lifting (hidden layer weights) learned via gradient-based meta-learning from the task-specific adaptation (output layer weights) solved via closed-form least squares. This design effectively decouples the optimization of the representation from the optimization of the solution, allowing for extremely fast inference and fine-tuning. The mathematical derivation linking the proposed linearization to the regularized Gauss-Newton method is sound and provides a strong theoretical foundation for the observed convergence rates.
Results & evaluation
The experimental evaluation is extensive, covering a wide range of 1D, 2D, and 3D nonlinear PDEs, including Burgers', KdV, Kuramoto-Sivashinsky, Navier-Stokes (Lid-Driven Cavity), and various reaction-diffusion equations. The paper demonstrates significant improvements over baselines like PINet (Picard-based), standard PINNs, and neural operators (DeepONet, FNO). Specifically, it claims up to three orders of magnitude lower relative error compared to recent neural solvers on a reaction-diffusion benchmark, while using 16x fewer training tasks and significantly less training time. The comparison with state-of-the-art meta-learning PINNs shows substantial speedups in fine-tuning time. The inclusion of both unsupervised (physics-only) and hybrid (data+physics) learning scenarios adds depth to the evaluation.
Reproducibility
The authors provide a public GitHub repository with code and datasets. The paper includes detailed descriptions of the model architecture, hyperparameters (meta-learned), and experimental setups. The use of standard libraries and clear algorithmic pseudocode enhances reproducibility. The specific details on temporal domain decomposition and initial guess strategies are well-documented in the appendix.
Limitations
The method relies on the construction of a dense matrix $A$ for the least-squares solve, which can become computationally expensive for very high-dimensional problems or large collocation sets, although the authors argue that meta-learning compact features mitigates this. The paper acknowledges that for chaotic systems like K-S, temporal decomposition is still required for long-time prediction, which limits the "single-shot" capability. Additionally, the method's performance on highly irregular geometries or complex multi-physics systems with strong coupling is not fully explored beyond the presented benchmarks.
Broader impact
This work has significant implications for scientific computing and engineering simulations. By enabling fast, data-efficient, and physics-consistent solutions to nonlinear PDEs, it can accelerate design optimization, real-time control, and surrogate modeling in fields such as fluid dynamics, climate science, and chemical engineering. The meta-learning aspect allows for rapid adaptation to new parameter regimes, which is crucial for applications where PDE parameters vary frequently. Newton-PINet introduces a novel meta-learning framework that combines Newton linearization with Tikhonov-regularized least-squares to achieve fast, data-efficient, and accurate solutions for nonlinear PDEs. By decoupling representation learning from task-specific adaptation and leveraging the superior convergence properties of Newton's method over Picard iteration, the paper demonstrates substantial improvements in both accuracy and computational efficiency compared to existing PINN and neural operator approaches, establishing a new benchmark for physics-informed meta-learning.
Author of this paper? Send a correction.