Source-linked AI summary
Physics-informed graph neural Galerkin networks: A unified framework for solving PDE-governed forward and inverse problems
Han Gao, Matthew J. Zahr, Jian-Xun Wang
TL;DR
The paper addresses scalability, boundary-enforcement, and geometry limitations in PINNs for forward and inverse PDE problems. It proposes a discrete framework combining graph convolutional networks with variational PDE structure, and reports effectiveness across computational-mechanics problems involving linear and nonlinear PDEs.
Problem
PINNs face high training costs, difficult strict boundary enforcement, and limited support for irregular geometries with unstructured meshes.
Method
The paper combines graph convolutional networks with a variational PDE formulation in a unified discrete PINN framework for forward and inverse problems.
Results
Numerical results demonstrate the framework’s effectiveness on several forward and inverse computational-mechanics problems governed by linear and nonlinear PDEs.
Takeaways & Limitations
The framework directly handles unstructured meshes, strictly imposes essential boundary conditions, and assimilates sparse observations without penalty-parameter tuning.
Takeaways & Limitations
Variational formulations face additional error from difficult quadrature construction, while essential boundary conditions cannot be hard-imposed in point-wise formulations.
Abstract
from arXiv · showhide
Despite the great promise of the physics-informed neural networks (PINNs) in solving forward and inverse problems, several technical challenges are present as roadblocks for more complex and realistic applications. First, most existing PINNs are based on point-wise formulation with fully-connected networks to learn continuous functions, which suffer from poor scalability and hard boundary enforcement. Second, the infinite search space over-complicates the non-convex optimization for network training. Third, although the convolutional neural network (CNN)-based discrete learning can significantly improve training efficiency, CNNs struggle to handle irregular geometries with unstructured meshes. To properly address these challenges, we present a novel discrete PINN framework based on graph convolutional network (GCN) and variational structure of PDE to solve forward and inverse partial differential equations (PDEs) in a unified manner. The use of a piecewise polynomial basis can reduce the dimension of search space and facilitate training and convergence. Without the need of tuning penalty parameters in classic PINNs, the proposed method can strictly impose boundary conditions and assimilate sparse data in both forward and inverse settings. The flexibility of GCNs is leveraged for irregular geometries with unstructured meshes. The effectiveness and merit of the proposed method are demonstrated over a variety of forward and inverse computational mechanics problems governed by both linear and nonlinear PDEs.
1. Introduction
The paper develops a discrete PINN framework combining graph convolutional networks with PDE variational structure to address training, boundary-enforcement, and geometry limitations in forward and inverse problems.
- Existing challenges: Continuous PINNs incur high training costs because point-wise formulations require extensive automatic-differentiation computations over many collocation points.These costs become problematic in high-dimensional spatiotemporal and parameter domains.
- Existing challenges: Strict enforcement of initial and boundary conditions remains difficult for continuous PINNs, particularly when labeled data are scarce or absent.Distance-based constructions have not demonstrated effectiveness on complex real-world geometries.
- Existing challenges: CNN-based discrete PINNs improve efficiency and scalability by learning solution fields directly and computing residual derivatives through numerical discretization.Existing CNN approaches generally use finite differences on regular Cartesian grids and therefore remain restricted for irregular domains.
- Proposed framework: The proposed framework applies generalized convolution directly to unstructured mesh data represented as graph-like discrete non-Euclidean manifolds.It uses finite-element-inspired graph convolution without rasterizing simplex or quadrilateral elements.
- Proposed framework: Its variational PDE formulation lowers differential-operator order through integration by parts and replaces many collocation points with fewer quadrature points.These choices reduce learning complexity and can lower training cost.
- Proposed framework: The method exactly satisfies essential boundary conditions without penalty-coefficient tuning and supports data assimilation for unified forward and inverse solutions.Numerical results cover computational-mechanics problems governed by linear and nonlinear PDEs.
2. Methodology
PI-GGN combines graph convolution, continuous Galerkin discretization, and piecewise polynomial bases to solve forward and inverse PDE problems on unstructured meshes. It enforces essential boundary conditions and assimilates observations through constrained constructions.
- Graph convolutional neural network for unstructured data: PI-GGN represents unstructured meshes as graphs and uses a GCN to learn nodal PDE solutions from mesh coordinates.Graph convolutions update nodal features through message passing over mesh connections.
- Variational PDE-informed loss function: The physics-informed loss is built from PDE residuals so conservation laws drive GCN training across steady-state PDEs including Poisson, elasticity, and Navier–Stokes equations.The formulation defines fluxes, source terms, and gradients within a generic steady-state PDE representation.
- Variational PDE-informed loss function: Piecewise polynomial bases and the continuous Galerkin method reduce the solution search space and reconstruct PDE residuals variationally.The basis is used to discretize conservation laws and formulate the physics-informed loss through weak residuals.
- Unifying forward and inverse solutions: The GCN outputs nodal solution vectors from coordinate graphs, while unknown PDE parameters can be trained jointly with network parameters.This construction supports both forward solution learning and inverse parameter estimation.
- Essential boundary conditions enforcement: Essential boundary conditions are satisfied automatically by construction rather than through penalty-based enforcement.Static condensation restricts optimization to unconstrained degrees of freedom while prescribed boundary values remain fixed.
- Unifying forward and inverse solutions: Observation data can be assimilated and unknown parameters or boundary conditions inferred without hyperparameter tuning by strictly constructing the GCN output.The framework solves PDE states and inverse quantities simultaneously under constrained optimization.
3. Numerical experiments
The numerical experiments evaluate PI-GGN on multiple computational mechanics PDEs in forward and inverse settings. They vary boundary-condition and parameter knowledge, compare sparse-data assimilation strategies, and retain one GCN architecture across cases.
- Experimental problems: PI-GGN is demonstrated on Poisson, linear elasticity, and Navier–Stokes equations with known or unknown boundary conditions and parameters.The experiments cover both forward and inverse computational mechanics problems.
- Sparse-data assimilation: Two sparse-observation assimilation strategies are compared to assess strict data enforcement for parameter and field inversion.The reported comparison focuses on the advantage of imposing observations strictly.
- Experimental setup: The same GCN architecture is used across all cases, with hidden-layer node dimensions [32, 64, 128, 256, 128, 64, 32].The relative error metric e is defined for evaluating the solutions.
3.1. Poisson equation
PI-GGN solves Poisson forward and inverse problems on square and circular domains, including unstructured geometries, while assimilating sparse observations to infer unknown sources and fields.
- Forward solution: On square and circular domains, PI-GGN predictions agree with FEM or analytical references, with reported errors e = 5 × 10−3 and e = 5 × 10−4.The square-domain result also reports e = 0.5% in the contour comparison.
- Forward solution: The graph-based discrete PINN handles non-rectangular domains with unstructured meshes without the rasterization or coordinate transformation required by standard FD-based CNNs.The method uses finite-element-style discretization and polynomial bases to represent the solution.
- Inverse solution: With sparse observations, PI-GGN simultaneously solves the diffusion field and infers an unknown source term in the inverse problem.The example assumes f = 2 is unknown and observes u at one point; both penalty and strict assimilation recover the source ground truth.
3.2. Linear elasticity equations
PI-GGN accurately solves forward linear-elasticity problems on regular, notched, and three-dimensional domains and infers unknown Lamé parameters from sparse displacement data.
- Forward solution: The unit-square forward displacement solution agrees very well with the FEM reference.The setup uses a 25-nodal graph with prescribed displacement and traction boundary conditions.
- Forward solution: The notched rectangular-domain solution is very accurate compared with FEM, with relative prediction error e = 1 × 10−2.The domain is discretized with simplex elements and requires no special treatment for irregular geometry.
- Forward solution: The three-dimensional hollow-cylinder displacement agrees reasonably well with FEM, although PI-GGN slightly over-predicts displacement at the right end.The model uses 40 hexahedral elements and 440 nodal points.
- Inverse solution: Using displacement observations at five randomly selected points, PI-GGN accurately infers the unknown Lamé parameters and reconstructs the entire displacement field.The true parameters are λ = µ = 1.
3.3. Naiver-Stokes equations
PI-GGN addresses nonlinear incompressible Navier–Stokes forward and inverse problems, producing accurate field predictions and more robust strict data assimilation than penalty-based inference.
- Forward solution: For lid-driven cavity flow, PI-GGN velocity and pressure contours agree well with FEM, with relative prediction errors below 1%.The test uses separate velocity and pressure representations in a mixed-element approximation.
- Forward solution: The same accuracy level used over 10000 collocation points for AD-based methods is achieved with the graph-based discretization.The comparison concerns the lid-driven cavity forward problem.
- Forward solution: In an idealized stenosis, both velocity and pressure fields are accurately solved and agree well with the FEM reference.The inlet velocity is prescribed at the bottom and a no-traction condition at the outlet.
- Inverse solution: For inverse inlet reconstruction, the relative error is e = 0.4 with soft penalty assimilation and e = 0.04 with hard enforcement.Sparse velocity observations are used to infer an inlet profile with more than 20 degrees of freedom.
- Inverse solution: Penalty-based inlet inference deteriorates despite using the same penalty coefficient, whereas strict assimilation avoids hyperparameter tuning and has better robustness.The penalty-based inferred inlet notably deviates from the ground truth.
4. Conclusion
The paper presents PI-GGN as a unified discrete PINN framework combining GCNs, Galerkin variational losses, finite-element discretization, and polynomial bases for forward and inverse PDEs.
- Framework: PI-GGN combines graph convolutional networks with Galerkin variational physics-informed loss functions to solve forward and inverse PDEs uniformly.The framework is presented as a discrete PINN formulation.
- Framework: The method naturally handles irregular domains with unstructured meshes while reducing the search space through polynomial representations.The reduced search space is associated with more efficient training.
- Data assimilation: Hard enforcement of boundary conditions and sparse observation data removes the need to tune penalty parameters and is reported to improve robustness.The conclusion links strict enforcement to both boundary conditions and data assimilation.
- Results: Numerical experiments across forward and inverse problems governed by linear and nonlinear PDEs demonstrate the effectiveness of the proposed method.The reported demonstrations include computational mechanics problems.
- Perspective: The authors position the framework as combining scientific deep learning with classic numerical techniques rather than isolating them.