mlsanityData to 7 Oct 2026

Does Text Steer Neural PDE Surrogates? A Controlled Diagnostic with OperatorCLIP

arXiv 2609.38517 ↗First posted 29 Sep 2026cs.LGcited by 0

Not on the map. Why?

Abstract (arXiv)

Lower error from a text-conditioned neural surrogate does not, by itself, show that the model uses the meaning of the text. We examine this attribution problem with OperatorCLIP, comparing an unconditioned FNO, a constant-sentence FiLM control, and a fixed task description trained with contrastive alignment. Three-seed experiments cover Darcy2D, ShallowWater2D, and three-dimensional compressible Navier-Stokes (CNS3D). Constant conditioning has lower mean test error on both 2D tasks. Relative to this control, task text plus alignment has a similar mean on ShallowWater2D and CNS3D and a higher mean on Darcy2D; these descriptive comparisons have substantial seed uncertainty. The latter comparison changes both prompt content and loss, so it isolates neither effect. The text encoder is trained from scratch, and each conditioned model sees only one description during training. In this regime, pairwise InfoNCE cannot identify matched pairs and has minimum log B. Prompt interventions show no reliable semantic ordering. This methodological caution demonstrates why pathway controls are needed; it neither establishes semantic competence of the encoder nor tests the effectiveness of text under varying physical context.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Builds on 28 of its 36 references that are in the map. First posted on arXiv 29 Sep 2026.

Builds on

28 of 36 references.

References · 28

  1. 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,456
  2. 2015U-Net: Convolutional Networks for Biomedical Image Segmentationcited by 19,604
  3. 2020Fourier Neural Operator for Parametric Partial Differential Equationscited by 2,806
  4. 2017FiLM: Visual Reasoning with a General Conditioning Layercited by 2,286
  5. 2019DeepONet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operatorscited by 1,965
  6. 2022Fourier Neural Operator with Learned Deformations for PDEs on General Geometriescited by 416
  7. 2022PDEBENCH: An Extensive Benchmark for Scientific Machine Learningcited by 396
  8. 2023GNOT: A General Neural Operator Transformer for Operator Learningcited by 321
  9. 2022Transformer for Partial Differential Equations' Operator Learningcited by 270
  10. 2023Convolutional Neural Operators for robust and accurate learning of PDEscited by 207
  11. 2022Towards Multi-spatiotemporal-scale Generalized PDE Modelingcited by 190
  12. 2024Poseidon: Efficient Foundation Models for PDEscited by 186
16 more
  1. 2024DPOT: Auto-Regressive Denoising Operator Transformer for Large-Scale PDE Pre-Trainingcited by 137
  2. 2024The Well: a Large-Scale Collection of Diverse Physics Simulations for Machine Learningcited by 134
  3. 2023In-Context Operator Learning with Data Prompts for Differential Equation Problemscited by 98
  4. 2024APEBench: A Benchmark for Autoregressive Neural Emulators of PDEscited by 51
  5. 2023Hypernetwork-based Meta-Learning for Low-Rank Physics-Informed Neural Networkscited by 42
  6. 2023CFDBench: A Large-Scale Benchmark for Machine Learning Methods in Fluid Dynamicscited by 40
  7. 2024PROSE-FD: A Multimodal PDE Foundation Model for Learning Multiple Operators for Forecasting Fluid Dynamicscited by 33
  8. 2024Unisolver: PDE-Conditional Transformers Towards Universal Neural PDE Solverscited by 31
  9. 2023GPT-PINN: Generative Pre-Trained Physics-Informed Neural Networks toward non-intrusive Meta-learning of parametric PDEscited by 20
  10. 2024OmniArch: Building Foundation Model For Scientific Computingcited by 8
  11. 2024Explain Like I'm Five: Using LLMs to Improve PDE Surrogate Models with Textcited by 6
  12. 2024PICL: Physics Informed Contrastive Learning for Partial Differential Equationscited by 5
  13. 2026Diagnosing Failure Modes of Neural Operators Across Diverse PDE Familiescited by 5
  14. 2026One Operator to Rule Them All? On Boundary-Indexed Operator Families in Neural PDE Solverscited by 5
  15. 2026ArGEnT: Arbitrary Geometry-encoded Transformer for Operator Learningcited by 3
  16. 2026Semigroup Consistency as a Diagnostic for Learned Physics Simulatorscited by 2

Built on by

Not cited yet.