A method-oriented collection on surrogate modeling and reduced-order methods — data-driven surrogates, POD/ROM reduction, operator learning and low-dimensional dynamics, with benchmark datasets and introductory surveys.
Updated Nov 3, 2026 67 resources
Introduction
Why this collection
Replacing a high-fidelity solver with an approximation that returns in seconds has moved from a paper trick to an engineering requirement. But behind the word "surrogate" sit four genuinely different technical routes, and picking the wrong one costs months.
Data-driven surrogates — fit the high-fidelity solver's output with a neural network; the goal is black-box accuracy and inference speed.
Reduced-order modeling (POD / ROM) — extract a low-dimensional basis from snapshots; stronger physical grounding and better interpretability.
Operator learning and low-dimensional dynamics — learn the solution operator directly, or use DMD / SINDy to extract dynamical modes, sidestepping the mesh.
Koopman and intrinsic dynamics — embed a nonlinear system into a linear manifold and study long-time behaviour spectrally.
This collection groups resources along those four routes, each with its reference tools, key papers and usable datasets. Decide first whether you need to accelerate a single solve or to describe the response surface of a parameterised family — then jump to the matching section and skip a lot of trial and error.
This collection complements the site's industry pages rather than duplicating them: industry pages answer "what is this industry using", while this one answers "what do these methods actually look like".
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
This paper investigates using artificial neural networks as surrogate models for black-box optimization, enabling efficient optimization when analytical function descriptions are unavailable by obtaining objective values through numerical simulation.
See LICENSE fileBlack-Box OptimizationSurrogate ModelNeural Network
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 a KAN-based single-parameterized surrogate model to accelerate drift kinetic equation solving for fast neoclassical toroidal viscosity torque calculation in tokamak plasma modeling.
See LICENSE fileTokamakPlasma PhysicsDrift Kinetic
This paper introduces a Bayesian surrogate-modeling framework for inferring regional Holzapfel-Ogden material parameters in a left ventricle partitioned into five physiological zones, using cardiac MRI data to overcome the limitation of treating the myocardium as mechanically homogeneous.
See LICENSE fileCardiac BiomechanicsBayesian InferenceSurrogate Modeling
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
Neural networks predict 0D reduced-order model parameters from vascular geometry, keeping the low cost and interpretability of the 0D model. Across aortic, aortofemoral and pulmonary patient-specific anatomies, mean pressure error drops by at least half, with the pulmonary cohort improving from 30.4% to 7.9%.
NewarXiv preprint (CC BY 4.0), copyright held by authorsHemodynamicsReduced OrderDigital Twin
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 36M-parameter UNet-based PDE solver v2 by Los Alamos National Laboratory, designed for surrogate modeling of partial differential equations in scientific computing.
A 36M-parameter neural network solver for partial differential equations based on U-Net architecture, enabling data-driven solving and prediction of PDE systems
The 500-variant DrivAer scale-resolving CFD dataset with 484 public runs, about 31 TB, official splits and geometry-extrapolation regimes, plus the surrogate regime it supports.
NewCC BY-SA 4.0 for the dataset; paper copyright retained by the authorsAutomotive AerodynamicsHigh-Fidelity CFDSurrogate Model
💬Editor note: A full industrial workflow combining an external-flow surrogate with wind-tunnel digitisation — useful as a deployment template.
The reservoir-simulation example shipped in NVIDIA PhysicsNeMo, where a single X-MeshGraphNet sub-example implements a multiscale graph neural surrogate for full-field pressure and water saturation prediction on ECLIPSE/IX-compatible output. The entry pins down the real input and target formats, how error accumulates over autoregressive rollout steps, and the current state of dataset availability, which the upstream README still marks as forthcoming.
NewApache-2.0 for the PhysicsNeMo example repository; the Norne dataset carries its own licence termsReservoir SimulationGraph Neural NetworkMeshGraphNet
💬Editor note: AI acceleration of reservoir simulation shows surrogates at energy-sector scale, and exposes the trade-offs under real compute constraints.
A physics-informed CNN whose loss encodes a finite-difference heat conduction residual, used as a surrogate for a liquid-cooled 21700 battery pack temperature field.
NewarXiv preprint (copyright held by authors)Thermal ManagementPINNBattery
A graph-neural-network medium-range weather forecasting model from Google DeepMind that outperforms the ECMWF operational deterministic system on 90 percent of 1380 verification targets and completes a 10-day global forecast in under a minute on a single accelerator, with code and weights now consolidated in the WeatherNext repository.
NewPaper is copyright of Science; code and notebooks are Apache-2.0, model weights permit commercial use since 2026-08-06, other materials are CC BY 4.0WeatherGraph Neural NetworkMedium-Range Forecasting
TopoDiff, the conditional diffusion approach to structural topology optimization from the MIT DeCoDe Lab, with its 33,000-sample dual-level test protocol, surrogate guidance and per-metric definitions. This entry verifies what the eightfold and elevenfold claims actually refer to, and records inference cost, the 3D extension and follow-up work.
NewCC BY 4.0 for the arXiv version; published at AAAI 2023Topology OptimizationDiffusion ModelGenerative Design
A 1000-case steady RANS airfoil benchmark with data-efficiency, Reynolds and angle-of-attack extrapolation tasks, reviewed for its numerical setup, metric definitions and unverifiable accuracy claims.
NewODbL v1.0 for dataset and code; paper copyright retained by the authorsAerospace AerodynamicsAirfoilHigh-Fidelity CFD
An open-source wave energy converter simulation framework led by NREL and Sandia, built on Simulink/Simscape Multibody with frequency-domain BEM hydrodynamics, convolution integrals and Morison elements, covering irregular waves, mooring and six-degree-of-freedom control, and providing reproducible time-domain ground truth for device design, control verification and surrogate model training.
NewApache-2.0 (Sandia LLC and NTESS, 2014, with certain rights retained by the U.S. Government under contract DE-NA0003525)Wave EnergyHydrodynamicsMultibody Simulation
A gradient-based aerodynamic shape optimization workflow built on SU2's discrete and continuous adjoint routes, covering airfoil, wing and automotive shape parameterizations, gradient verification and measured drag reduction, usable as a teacher signal for surrogate training and as a high-fidelity reference.
NewGNU Lesser General Public License v2.1Aerodynamic OptimizationAdjoint MethodCFD
An open-source seismic structural analysis framework originating at UC Berkeley, built on nonlinear fibre beams, shell elements and soil-structure interaction; through OpenSeesPy it supports scripted batch generation of earthquake response data for surrogate models and damage assessment.
NewCopyright The Regents of the University of California (1999-2020); free for non-commercial use by educational, research and non-profit entities, internal use only for other entities, commercial redistribution requires separate licensingOpenSeesEarthquake EngineeringNonlinear Analysis
A building-energy case built on EnergyPlus and its surrogate models, traced through the official documentation, the license file and the ORNL/UTK surrogate publications. The entry corrects the maintainer attribution, the release year and the way simulation fidelity and surrogate speed-up are usually quoted, then separates the quantitative claims a surrogate can actually support from the range where it stops being reliable.
NewEnergyPlus licence (BSD-style three-clause terms, name-use clause and retained U.S. Government rights)Building EnergyEnergyPlusHVAC
PreVAE-Turb applies a pre-trained variational autoencoder to spatial compression for surrogate modeling of gyrokinetic plasma turbulence. The model is fine-tuned with physics-informed losses, including a Fourier-space spectral loss to enforce spectral accuracy.
See LICENSE filePlasma TurbulenceSurrogate ModelingVariational Autoencoder
This paper develops a surrogate-assisted stochastic multiobjective framework for photovoltaic hosting-capacity planning, jointly optimizing penetration and total electrical loss under reliability constraints.
See LICENSE filephotovoltaic hosting capacitysurrogate modelingstochastic multiobjective optimization
DeepFEAv2 applies deep learning surrogate models to transient finite element analysis on unstructured meshes, enabling joint prediction of node-based and element-based outputs to reduce computational cost in structural mechanics.
See LICENSE fileFinite Element AnalysisDeep LearningTransient Simulation
A general scientific machine-learning method constructs dimensionally consistent surrogate models by deriving dimensionless groups and using truncated harmonic expansions to approximate their remaining dependence.
See LICENSE filedimensional consistencysurrogate modellingharmonic expansions
Proposes using ordinary multilayer perceptrons with finite-group averaging to achieve exact equivariant surrogate models for correlated electron systems, enabling efficient force and transition energy evaluations.
See LICENSE fileQuantum Many-BodyEquivariant Neural NetworksElectronic Structure
FIREWALL is a surrogate model for rapidly assessing localized runaway-electron wall loading and melting in tokamak plasma-facing components, supporting device-scale thermomechanical damage evaluation.
See LICENSE fileTokamakRunaway electronsPlasma wall loading
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
A global medium-range weather forecasting model published by Huawei Cloud in Nature in 2023, whose 0.25 degree forecasts outperform the ECMWF operational IFS across all tested variables, with a five-day forecast taking about 1.4 seconds; weights are CC BY-NC-SA 4.0 and non-commercial.
NewTrained weights are CC BY-NC-SA 4.0 with commercial use forbidden; the repository ships no LICENSE file and declares no code licenseWeather Forecasting3D TransformerMedium-Range Forecasting
PHASE uses reusable solution operators and residual error correction to model multiscale incompressible magnetohydrodynamics across varying parameter settings, supporting fast inference for plasma dynamics.
See LICENSE filemagnetohydrodynamicssurrogate modelingmulti-regime modeling
Proposes a nonlinear reduced-order model with convolutional operator compression to overcome the Kolmogorov N-width limitation of linear projection-based ROMs, achieving accurate approximations with fewer modes for parametrized PDEs.
See LICENSE filereduced-order modelconvolutional operatorN-width
💬Editor note: Worth noting: making the reduced-order model itself a differentiable operator, so it can be jointly optimised with the rest of the pipeline.
Proposes a ROM-based preconditioner to accelerate topology optimization, reducing computational cost of high-fidelity finite element analysis in structural conceptual design.
See LICENSE fileTopology OptimizationReduced Order ModelingPreconditioner
A comparative study benchmarking autoencoders against parsimonious diffusion maps (PDMs) grounded in classical numerical analysis for reduced-order modeling of Navier-Stokes turbulent flows.
See LICENSE fileNavier-StokesReduced-Order ModelingAutoencoder
💬Editor note: A direct comparison of autoencoder-based and numerics-inspired manifold learning — the best reference if you want to know exactly how autoencoder reduction differs from POD.
Proposes a ladder of information limits to quantify predictability boundaries for reduced-order models under limited information, covering memoryless and memory-augmented trajectory forecasts, event prediction, stationary statistics, and generative laws.
See LICENSE filereduced-order modelinformation theoryprediction theory
💬Editor note: An information-theoretic account of the predictability limits of reduced-order models — a useful prior for judging whether a problem can be reduced at all.
Introduces mollified-sharp decomposition for probabilistic regularization of moving shocks in parametric reduced-order models, enhancing shock-capturing capabilities in CFD applications.
See LICENSE fileshock capturereduced-order modelprobabilistic regularization
This paper develops a structure-informed data-driven ROM for transport-dominated systems governed by scalar hyperbolic conservation laws, using the kinetic defect formulation to separate characteristic transport from entropy defects localized on shock manifolds.
See LICENSE filereduced-order modelinghyperbolic conservation lawsshock capturing
A periodic physics-informed neural network (P-PINN) framework is proposed for solving parametric periodic flows and analyzing finite-amplitude aeroelastic responses in a purely physics-driven manner.
The paper presents a low-dimensional, invertible latent representation for reduced-order modeling of droplet morphology, enabling reconstruction of fluid-interface states from sparse recordings across a device’s actuation range.
See LICENSE fileDroplet morphologyFluid interfacesReduced-order modeling
This paper presents a two-stage reduction pipeline for parametrised PDE shape problems, combining Active Subspace-guided structured control-point selection with Sinkhorn autoencoders for free-form deformation and reduced-order modelling.
See LICENSE fileParametrised PDEsShape OptimisationFree-Form Deformation
A latent Lagrangian framework reduces the order of forced nonlinear dynamical systems by learning latent coordinates and neural representations of kinetic and potential energies, while force supervision removes the need for an ODE solver during training.
See LICENSE filenonlinear dynamical systemsreduced-order modelingphysics-informed machine learning
The paper introduces matrix-free moment-matching reduced-order modeling for multi-input quadratic-bilinear descriptor systems, avoiding explicit assembly and storage of full-order matrices and tensors.
See LICENSE filereduced-order modelingmoment matchingdescriptor systems
A Petrov-Galerkin operator inference method for stability-encouraging identification constructs reduced-order models directly from high-dimensional time-domain data for general physical-system modeling.
See LICENSE filePetrov-Galerkin methodoperator inferencereduced-order modeling
A reduced-order physical model describes re-ignition dynamics between consecutive discharges at the same position in dielectric barrier discharge arrangements, enabling time prediction without machine learning.
See LICENSE filedielectric barrier dischargeplasma physicsdischarge dynamics
This paper develops reduced-order modeling for mechanical metamaterials, addressing buckling and pattern transformation while accounting for local, nonlocal, and patterning fluctuation effects.
See LICENSE filemechanical metamaterialsreduced-order modelingcomputational homogenization
Uses patch-based autoencoder and neural ODE to build a distributed reduced-order model for studying relative periodic orbits in 2D Kolmogorov turbulence.
See LICENSE fileKolmogorov turbulenceperiodic orbitreduced-order model
SPOpInf is a non-intrusive method that learns stochastic differential equations from data for quadratically nonlinear systems with additive noise, supporting reduced-order modeling of complex stochastic dynamics and turbulent flows.
See LICENSE filestochastic operator inferencestochastic differential equationreduced-order modeling
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
PRONE introduces a Petrov-Galerkin regression framework Ψ(X)K≈Φ(Y) that unifies DMD, SINDy, Koopman regression, and sparse regression as variants of a single construction for data-driven dynamics learning.
See LICENSE fileOperator LearningPetrov-GalerkinDMD
💬Editor note: PRONE unifies DMD, SINDy and Koopman under a Petrov–Galerkin framework — read this if you want to understand how the three lines relate.
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
A scientific computing model on HuggingFace that integrates Dynamic Mode Decomposition with Neural Operator for scientific data modeling and prediction tasks.
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
Reveals compatibility conditions under which multi-symplectic Hamiltonian PDEs regularized by Brinkman-type penalization retain a multi-conformal symplectic structure, yielding exact local conservation laws for structure-preserving operator learning on complex geometries.
See LICENSE fileHamiltonian PDEsBrinkman penalizationStructure-preserving learning
This paper studies out-of-distribution regression in scientific machine learning, focusing on conditions for stable extrapolation in high-dimensional function and operator learning. It analyzes how assumptions on target functions or operators affect extrapolation beyond the training domain and reviews existing bounds based on distribution-discrepancy penalties.
See LICENSE filescientific machine learningout-of-distribution generalizationoperator learning
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
TM4FF introduces a physics-constrained operator learning framework combining Transformer attention with Mamba state-space models for accurate flow field prediction and denoising.
See LICENSE fileFlow Field PredictionOperator LearningTransformer
Large-scale CFD-generated wind field dataset with over 12,000 cases, designed for training and evaluating the terrain-informed machine learning model FuXi-CFD for fine-scale 3D wind field reconstruction.
See LICENSE filewind field simulationCFDmachine learning
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
💬Editor note: A dedicated surrogate-model benchmark with cross-PDE unified evaluation — the first choice when measuring generalisation.
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
Microsoft Research's open-source PDE benchmark suite with differentiable solvers and spatiotemporal dynamics data for evaluating long-horizon rollout stability.
A systematic survey of physics-knowledge + ML methods for PDE prediction, splitting the field into architectural-level (loss, structured prediction, augmentation) and data-level (multi-task, meta, contextual) approaches, with industrial perspective and open-source ecosystem review.
arXiv preprint (copyright held by authors)Physics-informed MLPDEMulti-task Learning
Systematic path from classical PDE numerics to data-driven physics AI - FEM, FVM, surrogates, PINN and their evolution.
CC BY 4.0RoadmapPDENumerical Method
💬Editor note: If you are unsure where the boundary between surrogates and PINNs lies, read this roadmap first — it lays out the whole technical spectrum.
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
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 — strictly a neural operator, but a foundational work for the "learn the operator" route, which shares its lineage with reduced-order modelling.
This survey reviews optimization of mathematical models combining integer decisions with ReLU neural networks for chemical and process systems engineering, particularly superstructure optimization using surrogates for complex nonlinear processes.
See LICENSE fileReLU neural networkssurrogate modelschemical process optimization