Neural Operators as Efficient Function Interpolators

Vasilis Niarchos, Angelos Sirbu, Sokratis TrifinopoulosarXiv 2026 · arXiv

TL;DR

This paper reframes Neural Operators (NOs), typically used for infinite-dimensional function spaces, as powerful tools for interpolating finite-dimensional functions. By viewing any function as an operator acting on an auxiliary base-space, NOs demonstrate competitive or superior performance to MLPs and KANs on various benchmarks, requiring fewer parameters and less training time.

What Problem Does This Paper Address?

Interpolating unknown functions from sparse data is a fundamental challenge across science and engineering. While traditional methods and standard neural networks (like MLPs) exist, they often struggle with high-dimensional or oscillatory functions and can be sensitive to data discretization. Kolmogorov-Arnold Networks (KANs) offer interpretability but can be computationally expensive. Neural Operators (NOs) are a recent advancement designed for learning maps between infinite-dimensional function spaces, particularly for solving partial differential equations (PDEs). However, their potential for the more common task of finite-dimensional function interpolation, especially in a way that leverages their unique properties like discretization invariance, has not been fully explored. This paper aims to bridge this gap by demonstrating how NOs can be effectively repurposed for this task.

Methods

The core innovation lies in a novel reframing of finite-dimensional function approximation as an operator learning problem. The authors introduce an "auxiliary base-space" and define an operator that acts by composition. Specifically, a target function $f: \mathbb{R}^{d_{in}} \to \mathbb{R}^{d_{out}}$ is transformed into an operator $F$ that maps functions from the auxiliary base-space $B$ to functions from $B$ to $\mathbb{R}^{d_{out}}$, where $F[x](s) = f(x(s))$. By learning this operator $F$ using a Neural Operator, the authors effectively learn the original function $f$.

This approach allows for different choices of the base-space dimension ($d_B$), enabling various interpolation strategies. The paper explores zero-dimensional (0D-NO), one-dimensional (1D-NO), and two-dimensional (2D) base-spaces. For the NO architecture, they primarily use Tensorized Fourier Neural Operators (TFNOs). The training process involves creating discretized input functions $x(s)$ from the training data, which are then fed to the NO. Inference also involves evaluating the learned operator on these constructed input functions.

Datasets

The paper utilizes several synthetic datasets for benchmarking:

  • Partial wave expansions: Functions of varying complexity and dimensionality.
  • Heaviside function: A function with a step discontinuity.
  • Piecewise combination of Gaussians and step functions: A more complex function combining features.
  • 1D discrete uniform noise: A challenging, unstructured function.
  • Hypergeometric functions: High-dimensional functions ($2F_1$ and $3F_2$) to test interpolation in higher dimensions.

For the real-world application, they use experimental binding energies from the Atomic Mass Evaluation 2020 (AME2020), specifically a subset of nuclei with $N > 12$ or $Z > 12$ and experimental uncertainties below 100 keV. This dataset is used to learn corrections to the Weizsäcker-Skyrme version-4 (WS4) mass model.

Key Insights and Ideas

  • NOs as powerful interpolators: The core insight is that NOs, designed for infinite-dimensional spaces, are surprisingly effective and efficient for interpolating finite-dimensional functions, often outperforming traditional MLPs and KANs.
  • Auxiliary base-space reframing: The key technical innovation is viewing function interpolation as operator learning via an auxiliary base-space, which naturally fits the NO architecture.
  • Discretization invariance advantage: The inherent discretization invariance of NOs (like FNOs and TFNOs) allows for zero-shot super-resolution and stable performance across different test resolutions, a significant advantage over methods sensitive to discretization.
  • Parameter and training efficiency: NOs achieve competitive accuracy with significantly fewer parameters and less training time compared to other state-of-the-art methods, making them highly practical.
  • "Non-local" learning: Unlike MLPs that learn point-wise, NOs learn in a "non-local" way across higher-dimensional subspaces, which appears to be a superior strategy for interpolation.

Key Results

  • Analytic Benchmarks:
    • 1D-TFNOs consistently ranked among the top performers, exhibiting stable RMSE across various test sizes and effectively learning high-frequency features and noise.
  • In many cases, 1D-TFNOs and 0D-TFNOs outperformed MLPs, with KANs also showing strong performance, particularly in specific regimes.
  • For hypergeometric functions (4D and 6D), 1D-TFNO demonstrated the lowest and most stable RMSE.
  • Nuclear Binding Energy Application:
    • A 30-member TFNO ensemble achieved a pooled five-fold out-of-fold (OOF) RMS error of 198.2 keV for predicting nuclear mass residuals, significantly reducing the raw WS4 model's error (282.5 keV).
  • This performance places the TFNO ensemble among the best recent neural network approaches for this task, while maintaining high parameter efficiency and short training times.
  • The TFNO model is small enough to train in minutes on a single GPU, and the ensemble can be trained in parallel with minimal engineering complexity.

Limitations and Future Work

  • Limitations:
    • The paper focuses on interpolation, and the performance of NOs for extrapolation is not explicitly discussed.
  • While KANs can be computationally expensive to train, the paper notes that their implementation in fastkan uses radial splines, which might limit their applicability for single-input KANs.
  • Direct comparison with other ML models in the nuclear physics literature can be delicate due to differing evaluation protocols and data subsets.
  • Future Work:
    • Combining NO interpolators with multi-task learning for predicting multiple nuclear observables simultaneously.
  • Investigating prospective extrapolation capabilities of NOs to newly measured or dripline nuclei.
  • Establishing formal approximation bounds for NOs in the function-interpolation setting.

Summary updated

Abstract

Neural operators (NOs) are designed to learn maps between infinite-dimensional function spaces. We propose a novel reframing of their use. By introducing an auxiliary base-space, any finite-dimensional function can be viewed as an operator acting by composition on functions of the base-space. Through a range of benchmarks on analytic functions of increasing complexity and dimensionality, we demonstrate that NOs can match or outperform standard multilayer perceptrons and Kolmogorov--Arnold Networks in accuracy while requiring significantly fewer parameters and training time. As a real-world application, we apply a two-dimensional Tensorized Fourier Neural Operator (TFNO) to the nuclear chart, learning a correction to state-of-the-art nuclear mass models as a partially observed residual field. A TFNO ensemble reaches a held-out root-mean-square error of 198.2 keV, placing it among the best recent neural-network approaches while retaining high parameter efficiency and short training times. More broadly, these results introduce NOs as a scalable framework for finite-dimensional function interpolation, from analytic benchmarks to structured scientific data.
BibTeX
@misc{niarchos2026neural,
  author = {Vasilis Niarchos and Angelos Sirbu and Sokratis Trifinopoulos},
  title = {{Neural Operators as Efficient Function Interpolators}},
  year = {2026},
  eprint = {2605.07792},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2605.07792},
}
Figure