A systematic map of the neural-operator family: from the DeepONet and FNO foundations through differentiable frameworks, physics-informed variants, datasets and benchmarks to real-world deployments.
Updated Nov 3, 2026 83 resources
Introduction
Why this collection
A PINN trains one network per equation to be solved. Neural operators take a different route: learning the solution operator itself. Train once, then solve a whole family of PDEs with different boundary conditions, parameters and geometries.
The payoff is direct: for a parameterised problem the response surface is trained once, and every new operating point becomes a forward pass. This is the most industrially valuable shape of surrogate modeling and reduced-order modeling.
The family has already split into several technical lines, and choosing between them requires seeing the differences clearly:
Branch structure — DeepONet's branch-trunk layout versus FNO's spectral layout, the latter being intrinsically mesh-independent.
How physics priors enter — writing conservation and dissipation into the architecture (GENERIC-FNO), or writing physical constraints into the loss.
Geometry handling — how an operator keeps its accuracy as geometry and mesh change.
Engineering reality — compute cost, memory footprint, and interfaces to existing solvers.
This collection groups resources along those four threads, from the foundational papers through to industrial deployments. If you are weighing a move from PINNs to neural operators, read the "Tutorials & Surveys" group first, then jump straight to "Method Frontiers & Variants".
Introduces neural operators with integral kernels parameterized in Fourier space, enabling resolution-invariant PDE solving - a foundational and highly cited work in physics AI.
arXiv preprint (copyright held by authors)FNONeural OperatorFoundational
💬Editor note: The FNO paper: mesh-independence comes from spectral parameterisation, which is why it remains the most widely adopted operator-learning baseline.
Presents a branch-trunk architecture realizing a deep learning version of the universal operator approximation theorem - the first systematic architecture for learning nonlinear operators and still a mainstream baseline.
arXiv preprint (copyright held by authors)DeepONetNeural OperatorFoundational
💬Editor note: The DeepONet paper: the branch-trunk layout opened the "learn the operator" route and, together with FNO, defines the two mainstream paradigms.
A fundamental physics neural operator model project published by Delta Lab on HuggingFace. Refer to official repository documentation for specific functionality and performance details.
A scientific computing model on HuggingFace that integrates Dynamic Mode Decomposition with Neural Operator for scientific data modeling and prediction tasks.
ETH Zurich CAM lab's PDE foundation model and downstream fine-tuning suite, using semigroup augmentation to mine abundant training pairs and cover 15 downstream tasks with mesh-invariant inference.
NewNo LICENSE file in the repository; the official model weights are labelled CC BY-NC 4.0 on HuggingFace and marked as research modelsFoundation ModelNeural OperatorDownstream Tasks
NVIDIA's physics-ML framework (formerly Modulus, renamed in 2025), a PyTorch-based library for PINN, neural operators and physics-constrained training - architecture, versions, comparisons and failure boundaries.
NewApache-2.0PINNNeural OperatorNVIDIA
💬Editor note: NVIDIA PhysicsNeMo turns neural operators into industrial-grade modules, with compute scheduling and deployment paths already worked out.
Reference PyTorch library for FNO, TFNO and DeepONet: function-space contracts, parameters, getting started, failure modes and cross-resolution evaluation.
NewMITNeural OperatorFNODeepONet
💬Editor note: The official NeuralOperator library, covering FNO, DeepONet, TNO and more — the first implementation to reach for when establishing a baseline.
VIOT is a generative neural operator for incompressible density transport that predicts divergence-free velocity fields via feed-forward inference, replacing hour-scale optimization in adjoint fluid solvers and enabling rapid transport trajectory generation.
See LICENSE fileIncompressible FlowNeural OperatorOptimal Transport
ICML 2026 paper proposing APT, a geometry-/mesh-/physics-agnostic neural operator that fuses graph-based local feature extraction with global attention for long-range physical effects; achieves 4× saturation error reduction on 2D faulted CO2 storage vs prior SOTA, 19,873× speedup over GEOS and 6,800× over ECLIPSE.
arXiv preprint (copyright held by authors)Neural OperatorSubsurface EnergyCO2 Storage
Proposes ADEx-FNO, a framework that extends Fourier neural operators to varying geometries and independent discretizations via ambient-domain embedding and signed distance functions, without modifying core Fourier-operator layers.
See LICENSE fileFourier neural operatorVarying geometry PDE solvingScientific machine learning
Proposes a collaborative framework of AI virtual labs under a citation-based influence economy to automate neural operator architecture search and training evaluation.
See LICENSE fileNeural OperatorAI AgentAutomated Scientific Discovery
The Anisotropic Maxwell Neural Operator enables rapid parametric modeling of ion cyclotron resonance heating full-wave responses for the EAST tokamak by learning the one-parameter dielectric-field family within a single operator.
See LICENSE filePlasma physicsICRHMaxwell equations
Proposes Adaptive Operator Transformation (AOT) to address structural diversity challenges in multi-PDE pre-training by adapting target solution operators rather than increasing model capacity, improving neural operator generalization across equations.
See LICENSE fileNeural OperatorPDE Pre-trainingAdaptive Transformation
A continually learning neural-operator surrogate model replaces the traditional forward solver for efficient Bayesian inversion of three-dimensional time-domain airborne electromagnetic data, addressing prohibitive computational costs.
See LICENSE fileAirborne EMBayesian InversionNeural Operator
Proposes an improved Fourier Neural Operator for modeling 2D Rayleigh-Bénard convection by predicting time increments, achieving a compact model (314k parameters, 7ms inference) while revealing accuracy limitations imposed by training data resolution.
See LICENSE fileFourier Neural OperatorRayleigh-Bénard ConvectionFluid Mechanics
Proposes the Fractional Fourier Neural Operator (FrFNO) that embeds the analytic Mittag-Leffler propagator to address the computational challenge of memory-dependent time derivatives and nonlocal fractional Laplacians in fractional PDEs, enabling efficient surrogate modeling.
See LICENSE fileFractional PDENeural OperatorMittag-Leffler Function
GENERIC-FNO embeds the metriplectic degeneracy structure of nonequilibrium thermodynamics into Fourier neural operators, coupling reversible energy-conserving and irreversible entropy-producing dynamics via degeneracy conditions, extending structure-preserving neural operators beyond single conservation laws.
See LICENSE fileneural operatorFourier neural operatorthermodynamics
💬Editor note: GENERIC-FNO encodes energy conservation and entropy production into the architecture — essential reading on the "physics in the structure, not the loss" route.
Proposes a geometry-aware anisotropic boundary correction method using neural operators as an efficient alternative to traditional CFD solvers for aerodynamic simulation, improving physical modeling near solid boundaries.
See LICENSE fileaerodynamicsboundary correctionneural operator
Proposes a geometry-aware incremental neural operator to address error accumulation in long-horizon autoregressive PDE prediction, mitigating spectral inconsistency and unstable channel couplings.
See LICENSE fileNeural OperatorPDE PredictionLong-Horizon
💬Editor note: Incremental neural operators address error accumulation in long-horizon rollouts — the key design for using operators in multi-step evolution.
HIPNO leverages symmetry-aware physics-informed neural operators to infer hemodynamic states from non-invasive signals, addressing scale symmetry challenges in physics-informed hemodynamic inference and expanding access to advanced monitoring.
See LICENSE fileHemodynamicsPhysics-Informed Neural NetworksNeural Operators
Proposes an adaptive memory gating mechanism enabling neural operators to dynamically adjust memory weights based on observation conditions, improving adaptability in solving time-dependent PDEs.
See LICENSE fileNeural OperatorsAdaptive MemoryTime-dependent PDE
💬Editor note: An adaptive memory-gating mechanism targeting drift in long-horizon prediction, with modest compute cost and a light retrofit.
A joint training approach that addresses neural operator out-of-distribution degradation and PINN multi-solution convergence errors by leveraging imperfect neural operators to guide PINN optimization.
Proposes Graph Wavelet Compressed Sensing (GWCS) for offline compression of graph signals via sparse wavelet-domain representations, reducing data preparation costs for scientific machine learning training.
See LICENSE filegraph signal processingwavelet transformcompressed sensing
NeuIDO proposes a neural intrinsic dynamics operator that internalizes physical constraints into world models for predicting 3D object interaction trajectories, bridging the gap between physics-informed 4D generation and true world models.
See LICENSE fileWorld ModelPhysics-Informed GenerationNeural Operator
Proposes a neural operator-based surrogate model for CFD-level transient thermal-hydraulic simulation of helical coil steam generators in SMRs, enabling real-time digital twin applications.
A 36M-parameter UNet-based PDE solver v2 by Los Alamos National Laboratory, designed for surrogate modeling of partial differential equations in scientific computing.
Proposes the Kuramoto Neural Operator, introducing coupled oscillator dynamics into operator learning to overcome limitations of fixed-basis architectures for physical phenomena governed by local interactions.
See LICENSE fileNeural OperatorPDE SolvingCoupled Oscillator
Therm-FM adapts a pretrained PDE foundation model to steady-state and transient 3D-IC thermal simulation, reducing data-generation costs for cross-design reuse.
See LICENSE file3D-IC thermal simulationNeural OperatorPDE foundation model
A Transformer-based neural operator approach for 3D wind field prediction over complex mountainous terrain, addressing bottlenecks in traditional CFD methods that require expert-intensive mesh generation and computationally expensive iterative solvers.
See LICENSE fileNeural OperatorsWind Field PredictionTransformer
Proposes WaveOp-LiteFM, a lightweight neural-operator flow matching approach that addresses the trade-off between computational cost and accuracy in satellite-to-radar precipitation retrieval for atmospheric remote sensing applications.
See LICENSE fileprecipitation retrievalneural operatorflow matching
ICLR 2026 paper proposing ATOM, a quasi-equivariant pretrained transformer neural operator that requires no explicit molecular graph and uses temporal attention for parallel decoding of multiple future states; curates TG80 (80 compounds, 2.5M fs trajectories) and demonstrates zero-shot generalization on MD17/RMD17/MD22.
arXiv preprint (copyright held by authors)Neural OperatorMolecular DynamicsTransformer
Reframes Transolver's Physics-Attention as a special case of linear attention and proposes LinearNO (Linear Attention Neural Operator); achieves SOTA on six PDE benchmarks with 40.0% fewer parameters and 36.2% less compute, plus superior performance on AirfRANS and Shape-Net Car industrial datasets.
arXiv preprint (copyright held by authors)Neural OperatorTransformerLinear Attention
A systematic review of 2025-2026 progress in physics-informed ML - neural operators, diffusion for PDE solving, large benchmarks and industrial deployment.
NVIDIA's PINN framework built on neural operators (FNO/DeepONet/AFNO) for end-to-end differentiable physics simulation, with multi-GPU training and deployment.
A physics-conditioned neural operator targets atrioventricular valve mechanics across varying pressure and tissue properties, aiming to accelerate repeated valve assessments compared with finite element analysis.
See LICENSE filePhysics-Conditioned Neural OperatorCardiac Valve MechanicsFinite Element Analysis
BAT-NO is a structural-mechanics surrogate for autoregressively predicting transient vehicle-component responses and crashworthiness under varying boundary conditions.
See LICENSE filevehicle crashworthinessneural operatorfinite element surrogate
A scientific machine learning framework adapts simulation-pretrained neural operators to real-world data while preserving useful pretrained structure through a shared repair module.
See LICENSE fileNeural operatorsPhysical simulationReal-world adaptation
This paper presents cluster attention neural operators for solving parametric partial differential equations. It targets physical-field simulation by learning solution operators to accelerate parameter-space mapping, while addressing the quadratic attention complexity associated with Transformer-based neural operators.
See LICENSE fileneural operatorsparametric PDEscluster attention
This paper trains neural operators for partial differential equation inverse problems, constrained optimization, design, and control by enforcing input-output sensitivity consistency with sketched tangent information generated on-the-fly.
See LICENSE fileneural operatorstangent consistencypartial differential equations
This study evaluates whether lower prediction errors translate into better physically relevant flow statistics for neural operators modeling flow around a NACA4418 airfoil, using paired CFD simulations and experimental particle-image velocimetry measurements.
See LICENSE fileairfoil flowneural operatorsphysics-grounded explanations
This paper provides an empirical and theoretical analysis of Transolver’s physics-attention mechanism for three-dimensional fluid dynamics, examining how unstructured mesh points are assigned to slices, tokens interact through self-attention, and results are broadcast back to the points.
See LICENSE fileneural operatorsphysics-attentionfluid dynamics
This paper investigates shared spatial computation for energy-efficient neural-operator virtual sensing. It evaluates operating-energy reductions in a heat-exchanger service while retaining the selected checkpoint and its evaluated predictions.
See LICENSE fileNeural OperatorsVirtual SensingEnergy-Efficient Computing
This paper proposes a physics-enhanced neural operator for turbine endwall film cooling by incorporating the superposition principle into a transformer-based model. The approach targets temperature-field prediction across cooling layouts with varying numbers of film holes, focusing on improved generalization in fluid cooling analysis.
See LICENSE fileFilm CoolingNeural OperatorTransformer
A function-space Transformer with adaptive anchors targets continuous-domain physical fields, addressing the trade-off between resolving localized variation and computation over fixed uniform grids. The abstract relates this approach to neural operators and latent-attention architectures.
See LICENSE filephysical fieldsneural operatorscontinuous-domain modeling
INO is a data-free neural operator for learning parametric PDE solution operators, trained directly on the PDE weak form without large simulation datasets or prolonged physics-informed GPU training.
See LICENSE fileneural operatordata-free learningPDE solution
KATOsuper presents an objective-agnostic framework that couples neural-reparameterized topology optimization with sensitivity-consistent Fourier neural operators, addressing gradient inconsistency between predicted objectives and sensitivities for stable structural optimization.
See LICENSE fileTopology OptimizationFourier Neural OperatorSensitivity Analysis
KernelOnet is an interpretable neural operator framework that incorporates kernel functions explicitly into its architecture, aligning the operator structure with boundary-type kernel-expansion forms for physics modeling.
See LICENSE fileneural operatorskernel functionsinterpretable machine learning
The paper presents a label-free PINN for the reliability of spatially variable slopes, combining finite-element virtual-work losses with incremental strength reduction for non-associated Mohr-Coulomb plasticity.
See LICENSE fileslope reliabilityphysics-informed neural networkneural operator
Proposes Interface Autoencoder (IAE) for learning neural operators on geometry-dependent PDEs, representing interface-function states in a fixed domain via finite projection codes.
See LICENSE fileNeural OperatorsPDEInterface Autoencoder
This paper studies zero-shot composition of neural operators for steady incompressible flow in idealized two-dimensional airway trees, auditing whether local consistency guarantees global conservation.
See LICENSE fileAirway FlowNeural OperatorsZero-Shot Composition
A high-performance PDE neural solver based on Manifold Function Encoder, featuring resolution-independent memory footprint and arbitrary PDE input support for large-scale physical simulations.
See LICENSE fileNeural OperatorPDE SolverMemory Efficiency
A physics-informed neural operator learns the map between function spaces in magnetohydrodynamic simulations, enabling efficient state inference for magnetic reconnection at any parameter and time step, significantly reducing computational cost at high Lundquist numbers.
See LICENSE fileMagnetohydrodynamicsMagnetic ReconnectionNeural Operator
This paper studies preconditioned training for physics-informed neural operators across general physics simulations, using governing equations rather than solver-generated datasets to reduce offline data-generation cost and enable fresh samples during optimization.
See LICENSE filephysics-informed neural operatorpreconditioned trainingscientific machine learning
PTNO trains a neural operator directly on noisy, low-cost Monte Carlo labels for particle transport in radiative transfer and plasma physics, reducing reliance on expensive converged simulations.
See LICENSE fileneural operatorparticle transportMonte Carlo simulation
This bioacoustics paper introduces a sampling-frequency-independent frontend that processes recordings at their native rates, combined with a Fourier Neural Operator backbone for heterogeneous multi-taxa classification while reducing high-frequency information loss from resampling.
See LICENSE filebioacousticsheterogeneous sampling ratesFourier Neural Operator
ScaleSplit-NO is a neural-operator method for three-dimensional turbulence prediction that exploits turbulence scale structure, reducing memory requirements and reliance on costly full-resolution training data in high-resolution simulation workflows.
See LICENSE fileNeural Operator3D TurbulenceTurbulence Prediction
SINO is a scale-invariant neural operator for partial differential equation solving, using low-rank parameterization to represent unresolved physics and closure mechanisms in coarse-grid computations.
See LICENSE filescale-invariant neural operatorpartial differential equationsphysics closure
The paper presents a structure-preserving Fourier Neural Operator for long-time Cahn-Hilliard phase-field dynamics using coarse temporal supervision, targeting efficient surrogate prediction during computationally intensive late-stage coarsening while addressing error accumulation in repeated time stepping.
See LICENSE fileCahn-Hilliard dynamicsFourier neural operatorphase-field simulation
This astrophysics paper uses neural-network emulation to model synchrotron self-Compton emission from very high-energy gamma-ray burst afterglows, while Bayesian inference addresses the computational cost of parameter estimation.
See LICENSE filegamma-ray burstshigh-energy afterglowsneural-network emulation
This paper studies rolling-element bearing fault diagnosis from vibration, a physical-sensing task, under operating-condition shift and a held-out-bearing protocol. It compares neural-operator and convolutional models while assigning each physical bearing entirely to either training or testing.
See LICENSE fileBearing Vibration DiagnosisDomain AdaptationNeural Operators
This paper evaluates physics-informed residual constraints in neural PDE surrogates for nonlinear dynamics, finding that known-equation residuals outperform the strongest matched generic regularizer under equal tuning budgets.
See LICENSE fileneural PDE surrogatesresidual constraintsnonlinear dynamics
A large-scale benchmark dataset for machine learning in computational fluid dynamics, covering four classic CFD problems—lid-driven cavity, laminar pipe flow, dam-break, and flow around a cylinder—to evaluate neural operator generalization under unseen conditions.
A PDE evaluation dataset containing 32 physical systems, each with four representations: source code, natural-language description, governing equation, and numerical trajectory, with both correct and corrupted variants.
See LICENSE filePDE evaluationcross-representation learningphysical systems
A benchmark dataset for PDE surrogate tasks within the ai4sci-tasks evaluation framework, containing training trajectories and evaluation splits for assessing neural operator and surrogate modeling performance.
See LICENSE filePDE SurrogateBenchmark DatasetScientific ML
Simulation training datasets for PDE control containing 3000 controlled trajectories generated by classical numerical solvers under randomly sampled control sequences, designed for training TI-DeepONet surrogate models.
See LICENSE filePDE controlDeepONetNeural operator
PhD course lecture notes from University of Bozen/Bolzano covering Physics Informed Neural Networks and Neural Operators with PyTorch implementation tutorials.
A free online textbook from TU Munich covering physics + deep learning: differentiable physics, PINNs, neural operators and turbulence modeling, with runnable code.
A practical focused review of neural operators covering DeepONet, FNO and PCANet with hands-on training strategies; benchmarked on Poisson and linear elasticity, and demonstrated as surrogates for Bayesian inversion - an engineering-ready primer.
arXiv preprint (copyright held by authors)Neural OperatorDeepONetFNO
💬Editor note: An engineering-perspective comparison: architectures benchmarked side by side on Poisson and linear-elasticity problems, plus demonstrations as Bayesian-inversion surrogates.
A systematic survey of operator learning covering the theory, architectures and error analysis of DeepONet, FNO and their variants - the reference entry point to learning infinite-dimensional maps.
arXiv preprint (copyright held by authors)Operator LearningSurveyDeepONet
💬Editor note: A systematic treatment of DeepONet and FNO — theory, architecture and error analysis; the best reference for infinite-dimensional mappings.
A survey of data-driven and physics-informed neural operators, organizing the architecture landscape (DeepONet / FNO / GNO / PINO / Transformer-based / diffusion) along three axes: scalability, generalizability, and data sensitivity.
arXiv preprint (copyright held by authors)Neural OperatorScientific MLDeepONet
NVIDIA open-source physics-AI training/inference framework (formerly Modulus) with official examples across CFD, structural, EM, weather, and semiconductor TCAD.
Semiconductor TCAD surrogates anchored on two verifiable foundations: the fp_ddm sub-example in NVIDIA PhysicsNeMo, which applies overlapping Schwarz domain decomposition to a steady-state 2D heat-conduction problem against a plane-stress elasticity case, and the etch-process morphology surrogates reported by the SK hynix TCAD Intelligence team. The entry separates process-level TCAD from device-level electrical surrogates and records the accuracy and geometry-scaling limits of each.
NewApache-2.0 for the PhysicsNeMo repository; DAC 2026 paper copyright held by the authorsSemiconductorTCADDomain Decomposition
A global data-driven weather forecasting model proposed by NVIDIA and NERSC in 2022 that produces a week-long forecast at 0.25-degree resolution in under two seconds, matching ECMWF IFS at short lead times for large-scale variables and beating it for fine-scale fields, while making thousand-member large ensembles an inexpensive routine operation.
NewCode is BSD 3-Clause (Copyright (c) 2022, FourCastNet authors); data follows the Copernicus CDS terms; model weights ship with the repository distribution notesWeather ForecastingFourier Neural OperatorDeep Learning
💬Editor note: FourCastNet is one of the few large-scale industrial deployments of neural operators, showing feasibility at global weather-data scale.