Neural Operator × Physics-Integration Knowledge Base

Two orthogonal axes · 7 × 6 = 42 cells · 120 primary-source papers · through June 2026

Cross-cutting concepts
Discretization Invariance Spectral Bias Universal Approximation of Operators Equivariance as Architectural Constraint Fixed-Point Iteration Foundation Model Pretraining for PDEs Linear vs. Nonlinear Operator Decomposition Attention ≈ Galerkin Projection (Transformer Legitimacy) The Heterogeneity Bottleneck for Foundation Models Three Competing Paradigms for Cross-PDE Generalization Open Frontiers and the Research Program
2019–2020

Foundational Architectures

Three architectures established operator learning as a distinct sub-field: DeepONet (branch-trunk inner product), FNO (Fourier-domain kernel), and GNO (graph message passing). Each proved a universal approximation theorem for operators and demonstrated zero-shot super-resolution on standard PDE benchmarks.

DeepONet: Learning Nonlinear Operators
arXiv:1910.03193 (2021)
Foundational branch-trunk architecture with universal approximation guarantee for operators.
Branch Net Trunk Net Universal Approximation Inner-Product Decomposition DeepONet
GNO: Graph Kernel Network for PDEs
arXiv:2003.03485 (ICLR 2020)
Graph message passing on point clouds; canonical architecture for irregular geometry.
Graph Message Passing Irregular Geometry Neighborhood Aggregation Edge Kernel GNO
MGNO: Multipole Graph NO for Long-Range Interactions
arXiv:2006.09535 (2020)
Hierarchical pooling extension to GNO for long-range interactions. See GNO card for keyword explanations.
FNO: Fourier Neural Operator for Parametric PDEs
arXiv:2010.08895 (ICLR 2021)
Fourier-domain kernel parameterization; O(N log N) complexity; discretization-invariant; the workhorse architecture.
Fourier-Domain Kernel Discretization Invariance Spectral Bias Universal Approximation FNO
Hidden Fluid Mechanics: A Navier-Stokes Informed Deep Learning Framework
arXiv:1808.04327 (Science 2020)
The foundational PINN paper for fluid mechanics: Navier-Stokes-informed neural networks recover hidden velocity/pressure fields from passive-scalar observations. Conceptual bridge from PINN to PINO.
Physics-Informed Neural Network Passive Scalar Transport Hidden Variables Navier-Stokes Residual HFM
2021–2022

Physics Integration

The PINO loss combines supervised data with PDE residuals at multiple resolutions; FourCastNet demonstrates FNO pretraining at climate scale. Physics-informed and foundation-model paradigms begin their rise.

PINO: Physics-Informed Neural Operator
arXiv:2111.03794 (JDS 2024)
Multi-resolution PDE residual loss on FNO; the canonical PINO recipe.
PDE Residual Loss Multi-Resolution Enforcement Soft Constraint Discretization Invariance PINO
FourCastNet: AFNO-Based Weather Foundation Model
arXiv:2202.11214 (2022)
Patch-level adaptive FNO; pretrained on ERA5; demonstrates NO foundation models can match traditional NWP.
AFNO ERA5 Pretraining Patch-Level Adaptive Fourier Weather Foundation Model FourCastNet
Neural Operator: Learning Maps Between Function Spaces
arXiv:2108.08481 (JMLR 2023)
Comprehensive theoretical framework unifying DeepONet, FNO, GNO under one operator-learning formalism. See DeepONet/FNO/GNO cards for keyword explanations.
Self-Attention for Operator Learning
arXiv:2105.14995 (2021)
First self-attention mechanism applied to PDE operator learning; linear-attention variant without softmax. Conceptual precursor to Transformer-based NOs (cell I.4).
Modular Neural Solvers for General PDE Systems
arXiv:2202.03376 (2022)
End-to-end learning across varying geometry, resolution, and boundary conditions via modular autoregressive neural PDE solvers. Hybrid (I.7) lineage.
MIONet: Multi-Input Operator Learning
arXiv:2202.06137 (SIAM J Sci Comput 2023)
DeepONet generalization to multiple input functions (initial condition + boundary + source term); universal approximation theorem for multi-input operators. See DeepONet card for keyword explanations.
2023

Equivariance and Geometric Structure

Hard physics-as-architecture emerges: G-FNO embeds group equivariance, PI-GANO uses signed-distance-field embeddings for cross-geometry generalization. Theoretical rate results arrive.

G-FNO: Group Equivariant FNO
arXiv:2306.05697 (ICML 2023)
FNO with hard group equivariance; the foundation of the 2026 equivariance surge.
Group Equivariance Hard vs. Soft Constraints Group Action Equivariant Fourier Layer G-FNO
High-Dimensional PDE Operator Learning
arXiv:2301.12664 (2023)
Efficient approximation of high-dimensional input-output PDE mappings; addresses the curse of dimensionality for coupled PDE systems.
Geometry-Aware Neural Operators for Irregular Meshes
arXiv:2302.14376 (2023)
Unified NO for irregular mesh, multi-input functions, and multi-scale phenomena. GNO+attention hybrid (I.7) for boundary shapes and parameter vectors.
ICON: In-Context Operator Networks
arXiv:2304.07993 (ICLR 2024)
Single NN as operator learner via in-context examples; eliminates retraining/fine-tuning for new PDEs. Foundation (II.6) lineage.
Transfer Learning for PDE Foundation Models
arXiv:2306.00258 (2023)
Characterizes transfer-learning behavior for pre-trained PDE solvers; validates the pretrain-then-finetune paradigm in scientific ML. Foundation (II.6) lineage.
2024

Transformers and Foundation Models

Attention-based NOs compete with FNO on irregular geometry; PI-GANO formalizes "geometry as architecture"; CoDA-NO introduces codomain pretraining for foundation models.

Transolver: A Fast Transformer Solver for PDEs
arXiv:2402.02366 (ICML 2024 Spotlight)
Physics-Attention with learned physics slices; 22% error reduction over FNO on standard benchmarks.
Physics-Attention Irregular Geometry Adaptive Slicing Adaptive Cost Transolver
CoDA-NO: Codomain Attention Neural Operator
arXiv:2403.12553 (NeurIPS 2024)
Codomain tokenization + self-supervised pretraining; >36% few-shot improvement on multiphysics PDEs.
Codomain Tokenization Self-Supervised Pretraining Few-Shot Fine-Tuning Multiphysics Pretraining CoDA-NO
DPOT: Auto-Regressive Denoising Operator Transformer
arXiv:2403.03542 (NeurIPS 2024)
Fourier-Attention hybrid backbone + auto-regressive denoising pretraining; up to 52% error reduction; scales to 0.5B params; canonical I.7×II.6 denoising-pretraining reference.
Fourier Attention Denoising Pretraining Balanced Data Sampling PDE Scaling Laws DPOT
PI-GANO: Physics-Informed Geometry-Aware NO
arXiv:2408.01600 (2024)
DCON backbone + SDF geometry embedding + PINO loss; cross-geometry error <3%.
SDF Embedding DCON Backbone Physics-Informed SDF Loss Cross-Geometry Generalization PI-GANO
DiffFNO: Diffusion Fourier Neural Operator
arXiv:2411.09911 (CVPR 2025)
FNO + score-based diffusion in function space; arbitrary-scale super-resolution with calibrated uncertainty.
Score-Based Diffusion Function-Space Diffusion Arbitrary-Scale Super-Resolution Smoothness Prior DiffFNO
MambaNO / MNO: State-Space Models vs. Transformers
arXiv:2410.02113 (JCP 2025)
Selective scan SSM as Axis I.6; O(N) linear complexity; up to 90% error reduction over Transformer baselines.
Selective Scan SSM Alias-Free Linear-Time Complexity Long-Range Memory MambaNO
Poseidon: Efficient Foundation Models for PDEs
arXiv:2405.19101 (NeurIPS 2024)
Multiscale operator transformer foundation model; time-conditioned layer normalization; pretrains on a few PDEs and generalizes to unseen ones with strong sample efficiency. Foundation (II.6) exemplar.
2025

Hybrid Composition and Iterative Refinement

Hybrid compositions become the SOTA for irregular geometry; IRNO formalizes FNO iteration as fixed-point solving; GAOT (NeurIPS 2025) demonstrates graph + transformer + spectral hybrids on industrial CFD.

NOWS: Neural Operator Warm Starts for Iterative Solvers
arXiv:2511.02481 (2025)
FNO as warm start for CG / multigrid; combines NO speed with classical-solver accuracy.
Warm Start Classical Solver Best of Both Worlds Domain-Specific NOWS
GAOT: Geometry Aware Operator Transformer
arXiv:2505.18781 (NeurIPS 2025)
MAGNO + ViT hybrid; SOTA on industrial 3D CFD; canonical I.7 reference.
MAGNO ViT Blocks Hybrid Composition Industrial CFD GAOT
Foundation Model for Continuum Dynamics
arXiv:2511.15684 (2025)
Cross-domain foundation model for continuum physics simulation; addresses data heterogeneity, long-rollout stability, and resolution differences. Foundation (II.6) lineage.
2026

Equivariance Surge and Foundation Frontiers

Two converging frontiers: (1) hard physics-as-architecture — GENERIC-FNO, EqGINO, gauge-equivariant NOs compose multiple symmetry groups; (2) foundation models for physics — CompNO, Therm-FM, MoE routing. The two converge at equivariant foundation models.

IRNO: Iterative Refinement NO as Learned Fixed-Point Solver
arXiv:2605.24041 (2026)
FNO iterated as fixed-point solver; principled spectral-bias mitigation with convergence theory; ERA5 16x weather super-resolution.
Fixed-Point Iteration Spectral Bias Mitigation Convergence Theory ERA5 16x Super-Resolution IRNO
GeoTransolver: Geometry-Aware Physics Attention Transformer
arXiv:2512.20399 (NVIDIA, Dec 2025)
Transolver backbone + GALE attention: physics-aware self-attention on state slices + cross-attention to multi-scale ball-query geometry and persistent global context. Industrial CAE on DrivAerML, SHIFT-SUV, SHIFT-Wing. Primary new evidence for I.4×II.3 (promoted from Special to Single).
GALE Attention Multi-Scale Ball Queries Persistent Geometry Conditioning PhysicsNeMo GeoTransolver
GeoPT: Scaling Physics Simulation via Lifted Geometric Pre-Training
arXiv:2602.20399 (Tsinghua/MIT, Feb 2026)
Transolver backbone + dynamics-lifted geometric pretraining on 1M+ ShapeNet samples. Random velocity field + transport trajectory = self-supervision. Reduces labeled-data requirements 20-60% on industrial CFD/CSR/radiosity. Opens the geometry-pretraining sub-paradigm of II.6.
Lifted Geometric Pretraining Synthetic Dynamics Web-Scale Geometry Industrial Transfer GeoPT
Neural Modular Physics for Elastic Simulation
arXiv:2512.15083 (Dec 2025)
Modular hybrid elastic simulator: neural constitutive module + neural integration module connected via intermediate physical quantities. Two-stage training (separate + joint) avoids module collapse. Canonical 2025 representative of the "modular hybrid" sub-line.
Modular Decomposition Neural Constitutive Module Neural Integration Module Two-Stage Modular Training NMP
AASM Benchmark: Applied Aerodynamics Surrogate Modeling Benchmark Cases
AIAA 2025-0036 (Bekemeyer, Hariharan, Wissink, Cornelius)
Four industrial-aerodynamics benchmark cases: airfoil aerodynamic database, missile stability/control, two surface-pressure datasets. AIAA AASM initiative; complements DrivAerML and CarBench as the 2025 industrial-CFD benchmark trio.
Discretization Invariance Airfoil Aerodynamic Database Missile Stability & Control Surface Pressure Datasets AASM
Survey papers and cross-cutting frontiers
2026 corpus
See the Open Frontiers and the Research Program cross-cutting concept for the 11 open cells and the 2026-2030 research program. CompNO (Compositional Foundation Model), Therm-FM (industrial foundation model), GENERIC-FNO (energy + entropy conservation), EqGINO (SE(3) + geometry), and gauge-equivariant NOs are the dominant 2026 directions.
Foundation Model Pretraining for PDEs Equivariance as Architectural Constraint Open Frontiers and the Research Program
Strong (3+ papers)
Single (1-2 papers)
Open (no papers)
Special