Spectral-Inspired Neural Operator Learning with Limited Data and Unknown Physics
Han Wan, Rui Zhang, Hao Sun · 2025
Read Spectral-Inspired Neural Operator Learning with Limited Data and Unknown Physics by Han Wan, Rui Zhang, Hao Sun (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-21573.txt.
Introduction
Simulating a physical system usually means knowing its equations, or drowning a neural network in data. Han Wan, Rui Zhang, and Hao Sun take a third road in a 2025 paper. Their Spectral-Inspired Neural Operator learns the dynamics from as few as two to five example trajectories, with the physics never stated.
The model works in frequency space, where it builds local and global spatial derivatives straight from frequency indices, assembling a compact stand-in for the hidden differential operator. A multiplicative block handles nonlinear effects, and a low-pass filter keeps the representation clean. On two- and three-dimensional equation benchmarks, the method beat rivals by one to two orders of magnitude in accuracy. Trained on five trajectories, it outperformed data-hungry methods trained on a thousand, and stayed predictive on out-of-distribution cases.
Catalog record
| CREATOR | Han Wan, Rui Zhang, Hao Sun |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 67 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-21573
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Spectral-Inspired Neural Operator Learning with Limited Data and Unknown Physics Han Wan, Rui Zhang, Hao Sun Abstract . Learning PDE dynamics from limited data with unknown physics is challenging. Existing neural PDE solvers either require large datasets or rely on known physics (e.g., PDE residuals or handcrafted stencils), leading to limited applicability. To address these challenges, we propose S pectral- I nspired N eural O perator (SINO), which can model complex systems from just 5 trajectories without requiring explicit PDE terms. Specifically, SINO automatically learns Fourier multipliers as functions of frequency indices, capturing both high-order derivatives and global/local couplings, thus enabling a compact representation of the underlying differential operators in physics-agnostic regimes. To model nonlinear effects, it employs a Π \Pi -block that performs multiplicative interactions on derivative features, followed by a spectral low-pass filter for de-aliasing. Extensive experiments on both 2D and 3D PDE benchmarks demonstrate that SINO achieves state-of-the-art performance, with improvements of 1-2 orders of magnitude in accuracy. Particularly, with only 5 training trajectories, SINO outperforms data-driven methods trained on 200 trajectories and remains predictive on challenging out-of-distribution cases where other methods fail. Conference: Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2; August 09–13, 2026; Jeju Island, Republic of Korea Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD ’26), August 09–13, 2026, Jeju Island, Republic of Korea DOI: 10.1145/3770855.3817831 ISBN: 979-8-4007-2259-2/2026/08 v2rtp1298 CCS: Computing methodologies Artificial intelligence CCS: Computing methodologies Modeling and simulation CCS: Applied computing Physics † † cc-license: by 1. Introduction Partial differential equations (PDEs) govern the evolution of numerous physical systems, from heat transfer ( Pletcher et al., 2012 ) and chemical reactions ( Ghergu and Radulescu, 2011 ) to fluid dynamics ( Holton and Hakim, 2013 ) . Accurately modeling and simulating such PDE-governed systems is essential for understanding natural phenomena and engineering applications. However, in many scientific computing scenarios, we face two fundamental challenges ( Azizzadenesheli et al., 2024 ; Zhang et al., 2023a ) : (i) only a few trajectories of data are available , and (ii) the governing physics is unknown . While traditional methods offer strong stability and interpretability, their reliance on a deep understanding of the physical system limits their application in scenarios where the underlying physics is unknown. In recent years, data-driven neural PDE solvers have been proposed to address the limitations of traditional numerical methods ( Azizzadenesheli et al., 2024 ) . Representative works include DeepONet ( Lu et al., 2021 ) , FNO ( Li et al., 2021 ) , and DPOT ( Hao et al., 2024 ) . By leveraging the expressive power of neural networks and large datasets, these methods learn mappings between function spaces. Therefore, they can reduce computational costs under coarse spatiotemporal resolutions and do not require explicit knowledge of the underlying PDEs. However, because they fit observed data with limited physical inductive biases, these methods are prone to overfitting the training distribution, resulting in high data demand and poor out-of-distribution generalization to unseen conditions. In many scientific and engineering domains, obtaining sufficient training data requires expensive physical experiments (e.g., wind tunnel testing or combustion experiments), which severely limits their practical applications ( Parente and Swaminathan, 2024 ; Li et al., 2024a ; Zhang et al., 2025b ) . To reduce the data demands of data-driven solvers, physics-aware methods incorporate physical knowledge into neural surrogates, yielding two main paradigms: physics-informed and physics-encoded ( Faroughi et al., 2022 ) . Physics - informed methods, such as PINNs ( Raissi et al., 2019 ) and PINO ( Li et al., 2024c ) , add soft constraints by including PDE residuals in the loss function. They require explicit forms of the governing PDEs and often suffer from training instability ( Krishnapriyan et al., 2021 ) . On the other hand, physics-encoded methods, like PeRCNN ( Rao et al., 2023 ; Rao et al., 2022 ) and TSM ( Sun et al., 2023 ) , embed physical rules into the network architecture, e.g., hardcoding finite difference stencils as convolutional kernels ( Rao et al., 2023 ) . When designing these models, one still needs complete or partial PDE terms as prior knowledge. Moreover, since many architectures borrow from local numerical schemes such as the finite difference method (FDM) or the finite volume method (FVM), they inherit limitations in handling high-order derivatives and global interactions ( McGreivy and Hakim, 2024 ) . Figure 1. Motivation of SINO. a. SINO targets operator learning in the limited-data, unknown-physics regime. b. Unlike physics-aware methods that combine a learnable local term with a known global operator, SINO learns the global/local interactions in Fourier domain via Freq2Vec. To address these challenges, we propose the S pectral- I nspired N eural O perator (SINO), which can accurately model complex systems from as few as 5 trajectories without any explicit PDE terms. The design of SINO is inspired by spectral methods, which are well known for their ability to capture global information with exponential accuracy ( Jingrun et al., 2022 ) . Specifically, the Frequency-to-Vector (Freq2Vec) module maps each frequency index into a learnable embedding, mimicking the derivative multipliers in spectral methods without relying on predefined operators. To further capture nonlinear dynamics, we employ a \Pi -block that models multiplicative interactions, together with a low-pass filter to mitigate aliasing effects. By inheriting the structure of spectral methods, SINO achieves stronger generalization from limited observations and demonstrates robust out-of-distribution (OOD) performance compared to data-driven approaches, while also handling global interactions and high-order derivatives more effectively than other physics-aware methods. Fig. 1 provides a positioning overview of SINO with data-driven and physics-aware methods. In summary, we make the following contributions: (1) Novel architecture . We propose SINO, a data-efficient neural operator for modeling complex spatiotemporal dynamics without requiring prior knowledge of PDE terms. We propose two key modules: Freq2Vec simulates Fourier multipliers as functions of frequency indices, and the nonlinear operator block captures complex interactions through efficient multiplicative operations. (2) State-of-the-art (SOTA) performance . SINO achieves SOTA performance across multiple 2D and 3D PDE benchmarks, delivering improvements of 1–2 orders of magnitude in accuracy compared to baselines. To our knowledge, SINO is among the first physics-aware methods capable of simulating globally coupled systems such as the NSEs without requiring explicit PDE terms. (3) Data efficiency and OOD handling . Due to the spectral-inspired design, SINO captures the underlying operator beyond memorizing training trajectories. With only 5 training trajectories, SINO outperforms data-driven methods trained on 200 trajectories and remains predictive in out-of-distribution (OOD) tests where other methods fail. 2. Related Work 2.1. Data-driven Neural PDE Solvers When provided with abundant data, data-driven methods learn mappings between function spaces, bypassing the need for explicit PDE formulation. Representative examples include the FNO ( Li et al., 2021 ) and its variants ( Tran et al., 2023 ; Wen et al., 2022 ) , DeepONet ( Lu et al., 2021 ) and its variants ( Venturi and Casey, 2023 ; Lee et al., 2023 ) . Transformer-based models ( Wu et al., 2024 ; Li et al., 2024b ) and graph-based architectures ( Pfaff et al., 2021 ; Brandstetter et al., 2022 ) extend neural PDE