Abstract:Recent advances in operator learning theory have improved our knowledge about learning maps between infinite dimensional spaces. However, for large-scale engineering problems such as concurrent multiscale simulation for mechanical properties, the training cost for the current operator learning methods is very high. The article presents a thorough analysis on the mathematical underpinnings of the operator learning paradigm and proposes a kernel learning method that maps between function spaces. We first provide a survey of modern kernel and operator learning theory, as well as discuss recent results and open problems. From there, the article presents an algorithm to how we can analytically approximate the piecewise constant functions on R for operator learning. This implies the potential feasibility of success of neural operators on clustered functions. Finally, a k-means clustered domain on the basis of a mechanistic response is considered and the Lippmann-Schwinger equation for micro-mechanical homogenization is solved. The article briefly discusses the mathematics of previous kernel learning methods and some preliminary results with those methods. The proposed kernel operator learning method uses graph kernel networks to come up with a mechanistic reduced order method for multiscale homogenization.

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is the problem of excessively high training costs of current operator learning methods in large - scale engineering problems (such as multi - scale simulation of mechanical properties). Specifically: 1. **High computational cost in large - scale engineering problems**: For large - scale engineering problems that need to handle a large number of degrees of freedom (for example, multi - scale simulation of mechanical properties), existing operator learning methods require extremely high computational resources during the training stage, which makes these methods impractical in practical applications. 2. **Limitations of existing methods**: Although traditional neural networks can effectively learn nonlinear mappings between finite - dimensional Euclidean spaces and have the universal approximation theorem, they have limitations when dealing with mappings between infinite - dimensional function spaces. In addition, when dealing with multi - scale material simulations, traditional methods usually rely on discretization methods, which not only increase computational complexity but also limit the generalization ability of the model. 3. **Challenges in multi - scale material simulations**: In multi - scale material simulations, although existing methods (such as hierarchical modeling and concurrent modeling) can capture the heterogeneity of material microstructures, they often need to make compromises between the accuracy of analyzing physical phenomena and the number of degrees of freedom to be solved. More accurate models usually need to solve a large number of degrees of freedom, resulting in excessively high computational costs. To solve these problems, the paper proposes a method based on kernel learning, aiming to learn mappings between function spaces through self - consistent clustering analysis (SCA). Specifically, the paper proposes the following innovations: - **Kernel learning method**: A new kernel learning method is introduced to map the relationships between function spaces. This method can not only reduce computational costs but also improve the generalization ability of the model. - **Graph Kernel Networks (GKN)**: Using graph - structured data, non - local interactions in material microstructures are captured by constructing graph kernel networks. Compared with traditional uniform grids or finite - element triangulations, graph kernel networks can better handle complex non - local characteristics. - **Fourier Neural Operators (FNO)**: Combined with Fourier transform, the computational efficiency is further improved by performing convolution operations in the frequency domain. FNO can not only quickly solve partial differential equations (PDEs) but also has the universal approximation theorem and is applicable to a wide range of nonlinear operators. In summary, the core objective of this paper is to solve the problems of high computational costs and limitations of existing operator learning methods in multi - scale material simulations by introducing new kernel learning methods and graph kernel networks, thereby providing more efficient and accurate solutions for large - scale engineering problems.

An Introduction to Kernel and Operator Learning Methods for Homogenization by Self-consistent Clustering Analysis

Generalized Convexity-Based Inexact Projection Method for Multiple Kernel Learning

Learning Homogenization for Elliptic Operators

Minimax Optimal Kernel Operator Learning via Multilevel Training

A Mathematical Analysis of Neural Operator Behaviors

Nonparametric learning of kernels in nonlocal operators

Bayesian deep operator learning for homogenized to fine-scale maps for multiscale PDE

Operator Learning: Algorithms and Analysis

HomoGenius: a Foundation Model of Homogenization for Rapid Prediction of Effective Mechanical Properties using Neural Operators

Neural Operator: Learning Maps Between Function Spaces

Thermodynamically-Informed Iterative Neural Operators for Heterogeneous Elastic Localization

A Kernel Machine Learning for Inverse Source and Scattering Problems

Multiple Operator-valued Kernel Learning

An Operator Learning Approach to Nonsmooth Optimal Control of Nonlinear PDEs

Learning non-stationary and discontinuous functions using clustering, classification and Gaussian process modelling

A Neural Network Approach for Homogenization of Multiscale Problems

A finite operator learning technique for mapping the elastic properties of microstructures to their mechanical deformations

Nonparametric Control-Koopman Operator Learning: Flexible and Scalable Models for Prediction and Control

Multiscale Neural Operator: Learning Fast and Grid-independent PDE Solvers

Consensus Local Graph for Multiple Kernel Clustering

Interfacing finite elements with deep neural operators for fast multiscale modeling of mechanics problems