Redefining Neural Operators in $d+1$ Dimensions for Embedding Evolution

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Song, Haoze, Li, Zhihao, Zhang, Xiaobo, Gan, Zecheng, Lai, Zhilu, Wang, Wei
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915760873406464
author Song, Haoze
Li, Zhihao
Zhang, Xiaobo
Gan, Zecheng
Lai, Zhilu
Wang, Wei
author_facet Song, Haoze
Li, Zhihao
Zhang, Xiaobo
Gan, Zecheng
Lai, Zhilu
Wang, Wei
contents Neural Operators (NOs) have emerged as powerful tools for learning mappings between function spaces. Among them, the kernel integral operator has been widely used in universally approximating architectures. Following the original formulation, most advancements focus on designing better parameterizations for the kernel over the original physical domain (with $d$ spatial dimensions, $d\in{1,2,3,\ldots}$). In contrast, embedding evolution remains largely unexplored, which often drives models toward brute-force embedding lengthening to improve approximation, but at the cost of substantially increased computation. In this paper, we introduce an auxiliary dimension that explicitly models embedding evolution in operator form, thereby redefining the NO framework in $d+1$ dimensions (the original $d$ dimensions plus one auxiliary dimension). Under this formulation, we develop a Schrödingerised Kernel Neural Operator (SKNO), which leverages Fourier-based operators to model the $d+1$ dimensional evolution. Across more than ten increasingly challenging benchmarks, ranging from the 1D heat equation to the highly nonlinear 3D Rayleigh-Taylor instability, SKNO consistently outperforms other baselines. We further validate its resolution invariance under mixed-resolution training and super-resolution inference, and evaluate zero-shot generalization to unseen temporal regimes. In addition, we present a broader set of design choices for the lifting and recovery operators, demonstrating their impact on SKNO's predictive performance.
format Preprint
id arxiv_https___arxiv_org_abs_2505_11766
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Redefining Neural Operators in $d+1$ Dimensions for Embedding Evolution
Song, Haoze
Li, Zhihao
Zhang, Xiaobo
Gan, Zecheng
Lai, Zhilu
Wang, Wei
Machine Learning
Artificial Intelligence
Quantum Physics
Neural Operators (NOs) have emerged as powerful tools for learning mappings between function spaces. Among them, the kernel integral operator has been widely used in universally approximating architectures. Following the original formulation, most advancements focus on designing better parameterizations for the kernel over the original physical domain (with $d$ spatial dimensions, $d\in{1,2,3,\ldots}$). In contrast, embedding evolution remains largely unexplored, which often drives models toward brute-force embedding lengthening to improve approximation, but at the cost of substantially increased computation. In this paper, we introduce an auxiliary dimension that explicitly models embedding evolution in operator form, thereby redefining the NO framework in $d+1$ dimensions (the original $d$ dimensions plus one auxiliary dimension). Under this formulation, we develop a Schrödingerised Kernel Neural Operator (SKNO), which leverages Fourier-based operators to model the $d+1$ dimensional evolution. Across more than ten increasingly challenging benchmarks, ranging from the 1D heat equation to the highly nonlinear 3D Rayleigh-Taylor instability, SKNO consistently outperforms other baselines. We further validate its resolution invariance under mixed-resolution training and super-resolution inference, and evaluate zero-shot generalization to unseen temporal regimes. In addition, we present a broader set of design choices for the lifting and recovery operators, demonstrating their impact on SKNO's predictive performance.
title Redefining Neural Operators in $d+1$ Dimensions for Embedding Evolution
topic Machine Learning
Artificial Intelligence
Quantum Physics
url https://arxiv.org/abs/2505.11766