Abstract:Recent work provides promising evidence that Physics-Informed Neural Networks (PINN) can efficiently solve partial differential equations (PDE). However, previous works have failed to provide guarantees on the worst-case residual error of a PINN across the spatio-temporal domain - a measure akin to the tolerance of numerical solvers - focusing instead on point-wise comparisons between their solution and the ones obtained by a solver on a set of inputs. In real-world applications, one cannot consider tests on a finite set of points to be sufficient grounds for deployment, as the performance could be substantially worse on a different set. To alleviate this issue, we establish guaranteed error-based conditions for PINNs over their continuous applicability domain. To verify the extent to which they hold, we introduce $\partial$-CROWN: a general, efficient and scalable post-training framework to bound PINN residual errors. We demonstrate its effectiveness in obtaining tight certificates by applying it to two classically studied PINNs - Burgers' and Schrödinger's equations -, and two more challenging ones with real-world applications - the Allan-Cahn and Diffusion-Sorption equations.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is that **Physics - Informed Neural Networks (PINN) lack a global upper - bound guarantee for the residual error when solving Partial Differential Equations (PDE)**. Specifically, although existing PINN research has demonstrated high accuracy at specific points, it has failed to provide a global upper - bound for the residual error across the entire space - time domain, which is an important drawback in practical applications. By introducing a framework named ∂ - CROWN, the paper aims to provide a theoretical guarantee for the global upper - bound of PINN's residual error, thereby enhancing the reliability and credibility of PINN in practical applications. ### Background of the Paper Physics - Informed Neural Networks (PINN) is a method that combines physical laws and deep learning to solve Partial Differential Equations (PDE). Although PINN has demonstrated its effectiveness in many fields, existing research has mainly focused on the accuracy of point estimates without providing a guarantee for global error. This limitation poses challenges for PINN in practical applications that require high reliability. ### Objectives of the Paper The main objectives of the paper are: 1. **Define the correctness conditions of PINN**: These conditions require that the solutions of PINN satisfy certain tolerances in terms of initial conditions, boundary conditions, and residual error. 2. **Develop an efficient error - certification framework**: Namely ∂ - CROWN, which is used to verify whether the residual error of PINN across the entire space - time domain satisfies the given tolerance. 3. **Demonstrate the effectiveness of the framework**: Verify the effectiveness of ∂ - CROWN through experiments on the classical Burgers equation, Schrödinger equation, and the Allan - Cahn equation and the diffusion - adsorption equation with practical application backgrounds. ### Methods of the Paper 1. **Define correctness conditions**: The paper defines three conditions to evaluate the correctness of PINN: - Upper - bound of initial - condition error: $\max_{\hat{x} \in D} |u_\theta(0, \hat{x}) - u_0(\hat{x})|^2 \leq \delta_0$ - Upper - bound of boundary - condition error: $\max_{t \in [0, T], \hat{x} \in \delta D} |au_\theta(x) + b\partial_n u_\theta(x) - u_b(x)|^2 \leq \delta_b$ - Upper - bound of residual - error: $\max_{x \in C} |f_\theta(x)|^2 \leq \epsilon$ 2. **Develop the ∂ - CROWN framework**: This framework is based on the CROWN method and calculates the upper - bound of PINN's residual error through linear relaxation techniques. Specific steps include: - Calculate the linear upper and lower bounds of PINN output and its partial derivatives. - Use the McCormick envelope to handle nonlinear terms and further calculate the upper and lower bounds of the residual error. - Refine the error bounds through the greedy input branching technique to improve their tightness. 3. **Experimental verification**: The paper conducts experiments on multiple classical PDEs to verify the effectiveness of the ∂ - CROWN framework and compares it with the traditional Monte Carlo method. ### Conclusion By introducing the ∂ - CROWN framework, the paper successfully provides a theoretical guarantee for the global upper - bound of PINN's residual error, filling the gap in existing research. Experimental results show that this framework can not only effectively calculate error bounds but also has high reliability and credibility in practical applications.

Efficient Error Certification for Physics-Informed Neural Networks

Recipes for when physics fails: recovering robust learning of physics informed neural networks

Convergence and error analysis of PINNs

A Physics Informed Neural Network for Time-Dependent Nonlinear and Higher Order Partial Differential Equations

Error-Aware B-PINNs: Improving Uncertainty Quantification in Bayesian Physics-Informed Neural Networks

Distributed physics informed neural network for data-efficient solution to partial differential equations

Physics Informed Neural Networks (PINNs) as intelligent computing technique for solving partial differential equations: Limitation and Future prospects

A Priori Error Estimation of Physics-Informed Neural Networks Solving Allen--Cahn and Cahn--Hilliard Equations

NeuralPDE: Automating Physics-Informed Neural Networks (PINNs) with Error Approximations

A Unified Framework for the Error Analysis of Physics-Informed Neural Networks

Residual-based error bound for physics-informed neural networks

Characterizing possible failure modes in physics-informed neural networks

Physics-informed attention-based neural network for solving non-linear partial differential equations

Estimates on the generalization error of Physics Informed Neural Networks (PINNs) for approximating a class of inverse problems for PDEs

A novel discretized physics-informed neural network model applied to the Navier-Stokes equations

A simple remedy for failure modes in physics informed neural networks

Data-Guided Physics-Informed Neural Networks for Solving Inverse Problems in Partial Differential Equations

Enforcing continuous symmetries in physics-informed neural network for solving forward and inverse problems of partial differential equations

Error Analysis of Physics-Informed Neural Networks for Approximating Dynamic PDEs of Second Order in Time

Understanding the Difficulty of Solving Cauchy Problems with PINNs

iPINNs: incremental learning for Physics-informed neural networks