Learning Geometric-Aware Quadrature Rules for Functional Minimization

Accurate numerical integration over non-uniform point clouds is a challenge for modern mesh-free machine learning solvers for partial differential equations (PDEs) using variational principles. While standard Monte Carlo (MC) methods are not capable of handling a non-uniform point cloud, modern neural network architectures can deal with permutation-invariant inputs, creating quadrature rules for any point cloud. In this work, we introduce QuadrANN, a Graph Neural Network (GNN) architecture designed to learn optimal quadrature weights directly from the underlying geometry of point clouds. The design of the model exploits a deep message-passing scheme where the initial layer encodes rich local geometric features from absolute and relative positions as well as an explicit local density measure. In contrast, the following layers incorporate a global context vector. These architectural choices allow the QuadrANN to generate a data-driven quadrature rule that is permutation-invariant and adaptive to both local point density and the overall domain shape. We test our methodology on a series of challenging test cases, including integration on convex and non-convex domains and estimating the solution of the Heat and Fokker-Planck equations. Across all the tests, QuadrANN reduces the variance of the integral estimation compared to standard Quasi-Monte Carlo methods by warping the point clouds to be more dense in critical areas where the integrands present certain singularities. This enhanced stability in critical areas of the domain at hand is critical for the optimization of energy functionals, leading to improved deep learning-based variational solvers.

💡 Research Summary

This paper, titled “Learning Geometric-Aware Quadrature Rules for Functional Minimization,” addresses a critical bottleneck in mesh-free machine learning solvers for partial differential equations (PDEs), such as the Deep Ritz Method and Physics-Informed Neural Networks. These solvers rely on the numerical integration of functionals over the domain, which is typically approximated using a weighted sum over a set of sample points. A significant challenge arises when these point clouds are non-uniform, adaptive, and unstructured—a common scenario in practice where sampling is influenced by domain geometry or physical phenomena. Standard quadrature rules (e.g., Gaussian) require grids, Monte Carlo methods converge slowly, and Kernel Optimal Quadrature is computationally prohibitive for dynamic point sets within a training loop.

To bridge this gap, the authors introduce QuadrANN, a novel Graph Neural Network (GNN) architecture designed to learn optimal quadrature weights directly and efficiently from the raw geometry of any given point cloud. The core innovation lies in QuadrANN’s ability to produce a permutation-invariant, geometry-aware quadrature rule. The model is not trained on specific PDE solutions but learns the fundamental property of integration exactness on a versatile basis set of test functions, comprising normalized Hermite polynomials (up to total degree 5) and randomized trigonometric terms.

The QuadrANN architecture is meticulously engineered to capture multi-scale geometric information. The initial layer encodes rich local features for each point by combining: (1) a high-frequency Positional Encoding of its coordinates, (2) relative positional information, and (3) an explicit local density estimate (the reciprocal of the average distance from a point’s neighbors to their local centroid). Subsequent message-passing layers incorporate a global context vector, ensuring that the weight assigned to each point is informed by the overall shape and structure of the entire domain (e.g., convex vs. non-convex). Inspired by DenseNets, features from all layers are concatenated for the final prediction, mitigating over-smoothing and preserving multi-scale information. A final softmax activation guarantees that the predicted weights are positive and sum to one, fulfilling the basic requirement of a quadrature rule.

For robust training and evaluation, the authors devise a method to generate challenging non-uniform point clouds. They start with a low-discrepancy Sobol’ sequence (quasi-random) and apply a deterministic, non-linear warping transformation. This process creates point sets with controllable density variations, simulating realistic adaptive sampling scenarios.

The proposed method is rigorously tested on a series of problems. First, on basic integration tasks over convex (unit hypercube) and non-convex (L-shaped) domains, QuadrANN consistently reduces the estimation variance compared to standard Quasi-Monte Carlo methods. Second, on point clouds warped to be denser near integrand singularities, QuadrANN successfully adapts its weights to the non-uniform density, demonstrating its geometric awareness. Finally, in practical applications, integrating QuadrANN into variational solvers for the Heat and Fokker-Planck equations leads to more accurate and stable solutions than those obtained using QMC integration.

In conclusion, QuadrANN presents a powerful, data-driven framework for learning accurate quadrature rules on fixed, non-uniform point clouds. By efficiently mapping geometry to optimal weights, it overcomes key computational limitations of prior methods and enhances the stability and accuracy of deep learning-based variational PDE solvers, marking a significant advance at the intersection of numerical analysis and machine learning.