Abstract:Data fields sampled on irregularly spaced points arise in many applications in the sciences and engineering. For regular grids, Convolutional Neural Networks (CNNs) have been successfully used to gaining benefits from weight sharing and invariances. We generalize CNNs by introducing methods for data on unstructured point clouds based on Generalized Moving Least Squares (GMLS). GMLS is a non-parametric technique for estimating linear bounded functionals from scattered data, and has recently been used in the literature for solving partial differential equations. By parameterizing the GMLS estimator, we obtain learning methods for operators with unstructured stencils. In GMLS-Nets the necessary calculations are local, readily parallelizable, and the estimator is supported by a rigorous approximation theory. We show how the framework may be used for unstructured physical data sets to perform functional regression to identify associated differential operators and to regress quantities of interest. The results suggest the architectures to be an attractive foundation for data-driven model development in scientific machine learning applications.

What problem does this paper attempt to address?

### What problem does this paper attempt to solve? This paper aims to solve the problem of dealing with **unstructured data** in scientific and engineering applications, especially those data with irregularly distributed sampling points. Specifically, the paper proposes a new framework named GMLS - Nets to address the following challenges: 1. **Limitations of CNNs**: - For data on regular grids, Convolutional Neural Networks (CNNs) have successfully utilized weight sharing and invariance to extract features. However, in many practical applications, data does not always exist in the form of regular grids but is distributed in irregular point clouds. Traditional CNNs are difficult to be directly applied to these unstructured data. 2. **Learning from unstructured data**: - GMLS - Nets solves the problem of how to learn on unstructured point - cloud data by introducing a method based on Generalized Moving Least Squares (GMLS). GMLS is a non - parametric technique for estimating linear bounded functionals from discrete data and has been used to solve Partial Differential Equations (PDEs). By parameterizing the GMLS estimator, GMLS - Nets can learn operators with unstructured templates. 3. **Applications in scientific machine learning**: - The paper shows how GMLS - Nets can be used in scientific machine learning (SciML), especially in performing functional regression from unstructured physical datasets to identify relevant differential operators and regress the quantities of interest. This provides a basis for constructing data - driven scientific models while retaining physical invariances such as conservation principles. 4. **Improving the performance of traditional scientific components**: - GMLS - Nets can also improve traditional scientific computing components, such as time integrators, and can regress the quantities of engineering interest from scientific simulation data. 5. **Handling complex geometries and fluid dynamics problems**: - The paper also explores the application of GMLS - Nets in handling complex geometries and fluid dynamics problems, for example, predicting drag through unstructured fluid velocity data without directly measuring pressure. ### Main contributions - Proposed the GMLS - Nets framework, which combines GMLS with deep learning and is suitable for unstructured data. - Demonstrated the effectiveness of GMLS - Nets in discovering partial differential equations, improving time integrators, and regressing engineering quantities from scientific simulation data. - Provided a new method to handle unstructured datasets, supporting data - driven modeling in scientific machine learning applications. Through these contributions, GMLS - Nets provides a powerful tool for dealing with unstructured data, especially in the fields of science and engineering.

GMLS-Nets: A framework for learning from unstructured data

MGCNN: a learnable multigrid solver for sparse linear systems from PDEs on structured grids

Neural Networks for Geospatial Data

DiffGCN: Graph Convolutional Networks via Differential Operators and Algebraic Multigrid Pooling

MgNet: A Unified Framework of Multigrid and Convolutional Neural Network

Generalized Moving Least-Squares for Solving Vector-valued PDEs on Unknown Manifolds

Flattening-Net: Deep Regular 2D Representation for 3D Point Cloud Analysis

RegGeoNet: Learning Regular Representations for Large-Scale 3D Point Clouds

A graph neural network-based framework to identify flow phenomena on unstructured meshes

Deep Convolutional Neural Network and Sparse Least Squares Migration

Scientific Computing Algorithms to Learn Enhanced Scalable Surrogates for Mesh Physics

LGSur‐Net: A Local Gaussian Surface Representation Network for Upsampling Highly Sparse Point Cloud

MMGP: a Mesh Morphing Gaussian Process-based machine learning method for regression of physical problems under non-parameterized geometrical variability

Geometric Neural Operators (GNPs) for Data-Driven Deep Learning of Non-Euclidean Operators

ManifoldNet: A Deep Network Framework for Manifold-valued Data

GPmap: A Unified Framework for Robotic Mapping Based on Sparse Gaussian Processes

A Tailored Convolutional Neural Network for Nonlinear Manifold Learning of Computational Physics Data using Unstructured Spatial Discretizations

NeuralMLS: Geometry-Aware Control Point Deformation

Deep Implicit Moving Least-Squares Functions for 3D Reconstruction

GS-Net: Generalizable Plug-and-Play 3D Gaussian Splatting Module

CS4ML: A general framework for active learning with arbitrary data based on Christoffel functions