QuadConv: Quadrature-Based Convolutions with Applications to Non-Uniform PDE Data Compression

Kevin Doherty,Cooper Simpson,Stephen Becker,Alireza Doostan

DOI: https://doi.org/10.1016/j.jcp.2023.112636

2023-08-28

Abstract:We present a new convolution layer for deep learning architectures which we call QuadConv -- an approximation to continuous convolution via quadrature. Our operator is developed explicitly for use on non-uniform, mesh-based data, and accomplishes this by learning a continuous kernel that can be sampled at arbitrary locations. Moreover, the construction of our operator admits an efficient implementation which we detail and construct. As an experimental validation of our operator, we consider the task of compressing partial differential equation (PDE) simulation data from fixed meshes. We show that QuadConv can match the performance of standard discrete convolutions on uniform grid data by comparing a QuadConv autoencoder (QCAE) to a standard convolutional autoencoder (CAE). Further, we show that the QCAE can maintain this accuracy even on non-uniform data. In both cases, QuadConv also outperforms alternative unstructured convolution methods such as graph convolution.

Machine Learning,Artificial Intelligence,Computational Engineering, Finance, and Science,Numerical Analysis

What problem does this paper attempt to address?

The paper is primarily dedicated to addressing the problem of convolution operations on non-uniform grid data, particularly in the application of partial differential equation (PDE) simulation data compression. Traditional discrete convolution operations assume that the input data is defined on a uniform grid, which becomes a limitation when dealing with non-uniform grid or point cloud data.

To solve this problem, the paper proposes a new convolution layer—QuadConv (convolution based on numerical integration), which is a method that approximates continuous convolution through numerical integration. QuadConv is specifically designed for non-uniform, grid-based data and achieves this by learning a continuous kernel function that can be sampled at arbitrary locations. Additionally, the paper details the efficient implementation of QuadConv.

To validate the effectiveness of QuadConv, the paper considers the task of compressing PDE simulation data on a fixed grid. Experimental results show that QuadConv can match the performance of standard discrete convolution on uniform grid data and maintain this accuracy even on non-uniform data. Furthermore, QuadConv outperforms other unstructured convolution methods, such as graph convolution.

In summary, the main contribution of this paper is the proposal of a novel convolution operation, QuadConv, suitable for non-uniform grid data, and the experimental validation of its effectiveness, especially in the application of PDE simulation data compression.