https://surrogatemodels.com/ News, papers, datasets, and funding and deals in Physics AI for engineers: Aero & CFD. en https://surrogatemodels.com/papers/tbnnhttps://surrogatemodels.com/papers/tbnnThu, 12 Nov 2026 09:00:00 GMTLearns the Reynolds stress anisotropy tensor of a RANS model from high-fidelity simulations. The network takes five invariants of the mean strain-rate and rotation-rate tensors and outputs the coefficients of ten basis tensors (Pope's integrity basis); a final multiplicative layer combines them, so the predicted tensor is Galilean invariant by construction.Aero & CFD https://surrogatemodels.com/papers/ledhttps://surrogatemodels.com/papers/ledThu, 12 Nov 2026 09:00:00 GMTAn autoencoder compresses the full simulation state to a few latent variables and an LSTM advances them in time (Latent-LED). Following the equation-free approach, the scheme can also alternate between the latent model and short runs of the original solver started from the decoded state, which corrects the accumulated error (Multiscale-LED).Aero & CFDGeneral PDEs https://surrogatemodels.com/papers/adaledhttps://surrogatemodels.com/papers/adaledThu, 12 Nov 2026 09:00:00 GMTCouples a running solver with a surrogate trained online during the same run: a convolutional autoencoder and an ensemble of probabilistic LSTMs. The simulation switches to the surrogate when its error against the solver and the ensemble uncertainty are below set thresholds, and returns to the solver when the uncertainty rises.Aero & CFDGeneral PDEs https://surrogatemodels.com/papers/scimarl-wall-modelshttps://surrogatemodels.com/papers/scimarl-wall-modelsThu, 12 Nov 2026 09:00:00 GMTLearns the wall model of a wall-modelled LES by multi-agent reinforcement learning. Agents along the wall read the velocity and its wall-normal gradient at one off-wall point and scale the local wall-shear stress by a factor between 0.9 and 1.1 at each step; the reward is the improvement in the predicted wall-shear stress. One variant (LLWM) takes instantaneous log-law coefficients as its state.Aero & CFD https://surrogatemodels.com/papers/marl-leshttps://surrogatemodels.com/papers/marl-lesThu, 12 Nov 2026 09:00:00 GMTTreats the subgrid-scale closure of a large-eddy simulation as a control policy. Agents spread over the grid observe invariants of the local velocity gradient and Hessian together with the global energy spectrum, and set the local Smagorinsky coefficient. One shared network is trained by reinforcement learning with the LES running in the loop, rewarded for matching the energy spectrum of DNS.Aero & CFD https://surrogatemodels.com/papers/ml-flow-initialisationhttps://surrogatemodels.com/papers/ml-flow-initialisationThu, 12 Nov 2026 09:00:00 GMTUses a DoMINO prediction as the initial field of an unsteady CFD run instead of uniform or potential flow: inside the region where the surrogate is trustworthy its field is used, and outside it is blended with uniform flow or a potential-flow solution. The solver itself is unchanged.Aero & CFD https://surrogatemodels.com/papers/graph-ledhttps://surrogatemodels.com/papers/graph-ledThu, 12 Nov 2026 09:00:00 GMTA graph neural network encoder reduces a flow field on an unstructured mesh to values on a smaller set of points, an autoregressive temporal attention model advances that reduced state, and a GNN decoder returns the field on the original mesh.Aero & CFD https://surrogatemodels.com/papers/simnethttps://surrogatemodels.com/papers/simnetThu, 12 Nov 2026 09:00:00 GMTA framework that trains physics-informed networks from the governing equations, with no simulation data, on geometry built from constructive solid geometry or STL. Design parameters such as fin heights are extra network inputs, so one training run covers a whole family of designs.Aero & CFDThermalBiomedical flows https://surrogatemodels.com/papers/clifford-layershttps://surrogatemodels.com/papers/clifford-layersThu, 12 Nov 2026 09:00:00 GMTScalar and vector fields at each point are packed into one multivector, and convolutions and Fourier layers act on these multivectors with Clifford algebra, so the coupling between, for example, pressure and velocity components is built into each layer.Aero & CFDElectromagneticsWeather & climate https://surrogatemodels.com/papers/g-ledhttps://surrogatemodels.com/papers/g-ledThu, 12 Nov 2026 09:00:00 GMTThe latent state is the simulation snapshot subsampled to a coarse grid, with no trained encoder. An autoregressive attention model advances it in time, and a diffusion model conditioned on the latent state generates the full-resolution fields.Aero & CFDGeneral PDEs https://surrogatemodels.com/papers/physicsbenchhttps://surrogatemodels.com/papers/physicsbenchMon, 09 Nov 2026 09:00:00 GMTA benchmark and leaderboard that ranks 66 models on nine datasets, expanded into 28 configurations over seven generation and prediction tasks in 1D, 2D and 3D, at data scales from S to XL, with a metric suite and a PageRank-based ranking (BenchRank) that debiases correlated metrics.Aero & CFDStructures & crash https://surrogatemodels.com/papers/acdmhttps://surrogatemodels.com/papers/acdmMon, 09 Nov 2026 09:00:00 GMTEach next flow state is generated by a conditional diffusion model from noise, conditioned on the previous states and the flow parameters, and the model is rolled out step by step; it is compared with U-Nets, ResNets, FNOs, latent transformers and PDE-Refiner.Aero & CFD https://surrogatemodels.com/papers/otnohttps://surrogatemodels.com/papers/otnoMon, 09 Nov 2026 09:00:00 GMTTreats each mesh as a density of points and solves an optimal transport problem that maps it onto a uniform density on a fixed latent grid, so every shape gets its own deformation. For a car, the surface mesh is mapped onto a 2D latent surface, a Fourier neural operator works there, and the result is mapped back to the mesh.Aero & CFD https://surrogatemodels.com/papers/deboshhttps://surrogatemodels.com/papers/deboshMon, 09 Nov 2026 09:00:00 GMTRuns Bayesian optimisation with a GNN surrogate instead of a Gaussian process. A 'reentrant' GNN predicts the pressure field and the performance number (drag or lift-to-drag), feeds the pressures back in and iterates; how fast the estimate converges is used as its uncertainty, which picks the next shapes to simulate.Aero & CFD https://surrogatemodels.com/papers/d-jepahttps://surrogatemodels.com/papers/d-jepaMon, 09 Nov 2026 09:00:00 GMTA JEPA-style encoder computes a 128-number latent from the geometry alone; the operating conditions enter only through small FiLM-modulated neural-field decoders, one per physics. A training loss forces the design parameters to be linearly readable from the latent, and case-level variance terms stop all geometries collapsing to the same code.Aero & CFDStructures & crash https://surrogatemodels.com/papers/globehttps://surrogatemodels.com/papers/globeMon, 09 Nov 2026 09:00:00 GMTBorrowed from boundary-element methods: the field at any point is a sum of learnable Green's-function-like kernels evaluated from every boundary face, with multi-scale branches, boundary-to-boundary communication layers and an explicit far-field decay. The model is translation, rotation, parity and units invariant.Aero & CFD https://surrogatemodels.com/papers/grounded-surrogatehttps://surrogatemodels.com/papers/grounded-surrogateMon, 09 Nov 2026 09:00:00 GMTA GeoTransolver surrogate trained on 2,300 CFD runs of the NASA Common Research Model wing-body is frozen; a small correction head on its per-cell latents is trained on wind-tunnel pressure-sensitive paint, registered onto the CFD surface, to learn the gap between CFD and the measurements.Aero & CFD https://surrogatemodels.com/papers/turtleshttps://surrogatemodels.com/papers/turtlesMon, 09 Nov 2026 09:00:00 GMTTurtLES is a 14 TB set of 3D turbulent cylinder wakes in the shear-layer transition regime from implicit large-eddy simulation: 400 trajectories of 400 time steps, 3-11 million points per frame, varying geometry, Reynolds number and angle of attack. Neural operators are benchmarked on autoregressive rollouts.Aero & CFD https://surrogatemodels.com/news/universalagi-suv-pthttps://surrogatemodels.com/news/universalagi-suv-ptMon, 02 Nov 2026 09:00:00 GMTUniversalAGI releases SUV-PT, built on its LIFT transformer architecture and trained, the company says, on millions of simulations from an automated pipeline. It predicts surface pressure, wall shear stress, volume pressure and velocity fields and drag, with a per-point uncertainty estimate. UniversalAGI reports drag-area MAPE of 1.1% on DrivAerML, 2.0% on its SUV-Bench Medium set and 16.8% (worst case 49.1%) on SUV-Bench Hard, which contains unrealistic mutated shapes. It is available on the company's platform and API.Aero & CFD https://surrogatemodels.com/news/autocfd5-ml-drivaermlhttps://surrogatemodels.com/news/autocfd5-ml-drivaermlWed, 28 Oct 2026 09:00:00 GMTAutoCFD5, the 5th Automotive CFD Prediction Workshop, took place on 8–9 October 2026 at DAMA Technopole in Bologna, organised by Upstream CFD with CINECA as local host. Its AI/ML Technology Focus Group, led by Neil Ashton (Mistral) and Astrid Walle (Pasteur Labs), used the open DrivAerML dataset as the standard training dataset to assess model architectures. Slides, recordings and group summaries were published on 28 October 2026. Talks in the AI/ML group came from Mistral, Cadence, IngeniAI, KIT, Luminary, Pasteur Labs and Siemens.Aero & CFD https://surrogatemodels.com/news/xplore-rwe-benchhttps://surrogatemodels.com/news/xplore-rwe-benchTue, 27 Oct 2026 09:00:00 GMTXplore builds what it calls high-fidelity reinforcement learning environments and evals around real engineering workflows. Its first public benchmark, Real World Engineering Bench (RWE-Bench), evaluates frontier AI models on mechanical tasks (parts and mechanisms modelled in CAD and checked for strength, fit and motion), electrical tasks (high-speed PCB routing, including DDR memory and PCIe links) and aerospace tasks (CFD and conjugate heat transfer, from building the mesh to a converged solution). Agents work in a sandbox with CadQuery, Gmsh, CalculiX and Python, and every run is graded by a deterministic verifier that runs physics solvers, with no LLM or human grading. Podcaster Dwarkesh Patel wrote that he is backing the company, founded by engineers Miguel and Carlos; the amount was not disclosed.Aero & CFDThermalElectromagnetics https://surrogatemodels.com/news/luminary-coupled-lpm-aerostructuralhttps://surrogatemodels.com/news/luminary-coupled-lpm-aerostructuralMon, 26 Oct 2026 09:00:00 GMTLuminary Cloud replaces the solvers in a static aeroelastic loop on a transonic swept wing. With SHIFT-Wing (pressure and shear on the surface) coupled to the TACS structural solver, a converged loop took 8 to 11 minutes on one NVIDIA T4 against about 4 hours for the CFD loop on three H100 GPUs; on three wings both converged in four iterations and the predicted lift reduction differed by 0.2 to 0.8 percentage points. A second model, SHIFT-Wingbox, trained on about 1,400 TACS solutions, then replaces the structure, and the two surrogates agree with the SHIFT-Wing and TACS loop within 1 percentage point of lift reduction and 0.04° of tip twist.Aero & CFDStructures & crash https://surrogatemodels.com/news/luminary-uq-conformal-error-barshttps://surrogatemodels.com/news/luminary-uq-conformal-error-barsTue, 06 Oct 2026 09:00:00 GMTLuminary Cloud's Aashwin Mishra and Dheeraj Vemula describe an uncertainty method for models built on the Luminary-SMART architecture. The trained model is frozen; a companion network of about 166,000 parameters, roughly 1% of the model, takes its per-point features, a geometry summary and the prediction, and returns an error-bar size for each field. It is trained with an interval-score loss on the model's errors, then calibrated on held-out simulations with normalized split conformal prediction; drag and lift are calibrated separately. The post uses 90% coverage and reports results on SHIFT-Wing, SHIFT-SUV, SHIFT-Truck and DrivAerML: on SHIFT-Wing the drag bar is ±0.0014 (14 counts) and contains the CFD drag for 69 of 73 test designs, and on SHIFT-SUV the bars contain the CFD value at 90.0% of surface points on 74 held-out designs. Luminary states that the calibration does not extend to designs far from the calibration set, which its out-of-distribution score is meant to flag first.Aero & CFD https://surrogatemodels.com/papers/cantohttps://surrogatemodels.com/news/cantoTue, 01 Sep 2026 09:00:00 GMTTurns each NURBS patch of the CAD model into one token from its control points, weights and knot vectors, with no meshing or point sampling of the geometry. A transformer combines the patch tokens into a geometry context that conditions an anchored neural-field decoder (taken from AB-UPT), which predicts surface and volume fields at any query point. The model is differentiable with respect to the CAD parameters.Aero & CFD https://surrogatemodels.com/news/luminary-ood-confidence-shift-winghttps://surrogatemodels.com/news/luminary-ood-confidence-shift-wingTue, 11 Aug 2026 09:00:00 GMTLuminary Cloud's Aashwin Mishra and Dheeraj Vemula describe Luminary-OOD, a detector that projects the incoming geometry and parameters into an embedding space and returns a confidence score before the physics model runs. Tested on the SHIFT-Wing dataset with the Luminary-SMART model, the mean score is 0.99 for in-distribution inputs, 0.92 for a slight shift in Mach number, 0.82 for a large one, 0.57 for the NASA CRM (a slight geometric shift, an extra tail section) and 0.13 for an SUV geometry. The post puts Monte Carlo dropout at about 50 times the inference cost and deep ensembles at about 5 times the training and inference cost, and calls its own overhead negligible, without a figure.Aero & CFD https://surrogatemodels.com/news/physicsx-px-car-aero-v1https://surrogatemodels.com/news/physicsx-px-car-aero-v1Fri, 07 Aug 2026 09:00:00 GMTPhysicsX makes PX-Car-Aero-v1 available on its platform: a model for external car aerodynamics pre-trained on PXNetCar, which now holds 33,203 RANS k-omega SST simulations from 444 baseline vehicles with a median of 88 morphed variants each. It predicts drag and lift coefficients, surface pressure and wall shear stress, and volume pressure and velocity. The post compares fine-tuning with training the same architecture from scratch. On a production SUV with a 21-parameter design space simulated with DES, fine-tuning needs roughly 84.4% fewer simulations to reach 85% directional accuracy on drag; on DrivAerML, with 5 training and 2 validation simulations, it reaches 90% against 70% from scratch. PhysicsX states that accuracy degrades on genuinely new designs on every metric. No weights, code or data are released.Aero & CFD https://surrogatemodels.com/news/beyondmath-physics-factory-drivaermlhttps://surrogatemodels.com/news/beyondmath-physics-factory-drivaermlMon, 03 Aug 2026 09:00:00 GMTBeyondMath says its Physics Factory agent crews analysed the DrivAerML data, then composed, trained and evaluated architectures from its building blocks, with under 30 minutes of human hands-on time. It reports relative L2 errors of 3.53% on surface pressure, 3.83% on volume pressure and 4.98% on velocity, and drag R² of 0.985 and lift R² of 0.975. In its table Transolver-3 is ahead on velocity (4.14%) and lift (0.985), and AB-UPT and Transolver trail on pressure.Aero & CFD https://surrogatemodels.com/news/physicsnemo-26-08https://surrogatemodels.com/news/physicsnemo-26-08Sat, 01 Aug 2026 09:00:00 GMTThe 26.08 release of NVIDIA PhysicsNeMo, the open-source framework for physics AI, moves the GeoTransolver and FLARE models out of the experimental namespace and adds experimental AeroJEPA, xDeepONet and Strata (weather emulation) models. On the data side it adds a Zarr backend, implicit-domain and exact-boundary mesh generation, and GPU remeshing with field transfer. The previous release, 26.05, focused on multi-dataset mesh training, data curation and active learning.Aero & CFDWeather & climate https://surrogatemodels.com/news/physicsnemo-cfd-car-aerohttps://surrogatemodels.com/news/physicsnemo-cfd-car-aeroFri, 29 May 2026 09:00:00 GMTNVIDIA's PhysicsNeMo-CFD module, released with PhysicsNeMo 26.05, scores pretrained models on DrivAerML or a user's own dataset with L2 errors on pressure, turbulent viscosity and velocity, and drag and lift coefficient errors. Pretrained checkpoints for GeoTransolver, DoMINO, MeshGraphNet and FigConvNet are published on Hugging Face. The post argues for shared datasets, metrics and leaderboards so that domain experts can test physical consistency and edge cases themselves.Aero & CFD https://surrogatemodels.com/news/physicsnemo-2-1https://surrogatemodels.com/news/physicsnemo-2-1Wed, 27 May 2026 09:00:00 GMTNVIDIA publishes PhysicsNeMo v2.1.0 on GitHub. A new unified external-aerodynamics recipe trains DoMINO, Transolver, GeoTransolver or FLARE on DrivAerML, SHIFT-SUV, HiLiftAeroML or a user's own data, mixing datasets at run time through mesh datasets and non-dimensionalisation, with surface or volume data. The release adds the GLOBE model with AirfRANS and DrivAerML examples, concrete-dropout uncertainty estimates and an out-of-distribution guard (OODGuard) for GeoTransolver, a `physicsnemo.sym` module that replaces the separate PhysicsNeMo-Sym package for PDE residuals, and a multi-diffusion subpackage for patch-based training and sampling. PyTorch 2.5 becomes the minimum version.Aero & CFDWeather & climate https://surrogatemodels.com/news/noether-2026-4https://surrogatemodels.com/news/noether-2026-4Wed, 29 Apr 2026 09:00:00 GMTEmmi AI releases Noether v2026.4.0, the first version under calendar-based numbering. AB-UPT no longer has a fixed surface and volume split: a `ModelDataSpecs` dictionary lets it take any number of named domains (gas, liquid, solid, multi-phase), and new untied physics blocks give each branch its own weights. New recipes cover DrivAerML end to end with query-based inference, the SIMSHIFT Heatsink benchmark (819 simulations; velocity, temperature and pressure on 3D volume fields, with AB-UPT or Transolver++) and a subset of the Emmi-Wing dataset on Hugging Face. Emmi says RoPE scaling and the Muon optimiser improve its AB-UPT baseline on DrivAerML; the release notes link a Weights & Biases report and give no figure. Muon support splits the attention weights, so earlier AB-UPT checkpoints do not load.Aero & CFDThermal https://surrogatemodels.com/news/luminary-shift-missilehttps://surrogatemodels.com/news/luminary-shift-missileTue, 07 Apr 2026 09:00:00 GMTLuminary Cloud introduces SHIFT-Missile, trained on about 6,000 high-fidelity RANS simulations of generic missile shapes from a parameterisation with more than thirty variables. It predicts surface pressure and wall shear stress, and derives integrated forces and moments, at Mach 1.5, 2.0, 2.5 and 3.0 and angles of attack of 0 to 4 degrees. On 2,500 test samples Luminary reports R² above 0.999 for axial force, normal force and pitching moment and MAPE below 1%. Predictions take seconds on a single consumer-grade GPU, according to the company.Aero & CFD https://surrogatemodels.com/news/physicsnemo-2-0https://surrogatemodels.com/news/physicsnemo-2-0Tue, 10 Mar 2026 09:00:00 GMTNVIDIA releases PhysicsNeMo v2.0.0, a refactor of the open-source framework that the 26.03 release notes describe as a major architectural change toward a modular toolkit. Reusable layers move to physicsnemo.nn, the base Module class to physicsnemo.core, and physicsnemo.launch is removed; a compatibility layer (physicsnemo.compat) keeps old import paths working, and a migration guide lists the changes. New packages are physicsnemo.diffusion (noise schedules, samplers, losses, multi-diffusion), physicsnemo.mesh (GPU mesh processing for simplicial meshes of any dimension) and refactored GPU datapipes. The 26.03 notes also add training recipes for crash dynamics with Transolver and MeshGraphNet, for car aerodynamics volume fields with GeoTransolver, and for StormScope with a diffusion transformer.Aero & CFDStructures & crashWeather & climate https://surrogatemodels.com/news/physicsx-pxnetcarhttps://surrogatemodels.com/news/physicsx-pxnetcarTue, 03 Mar 2026 09:00:00 GMTPhysicsX argues that DrivAerNet++ (8,121 geometries from three baselines) and Luminary SHIFT-SUV (1,994 geometries from two baselines) vary too little to test generalisation, because random splits put morphs of the same baseline in both training and test sets. It built PXNetCar, a proprietary dataset of 20,486 RANS simulations in STAR-CCM+ over 260 baseline vehicles with about 100 variants each. PhysicsX reports that state-of-the-art architectures generalise poorly to new vehicles and across CFD setups, that out-of-distribution error falls steadily as training data becomes more diverse, and that training on several datasets recovers in-simulator accuracy. The dataset is not released.Aero & CFD https://surrogatemodels.com/market/beyondmath-seedhttps://surrogatemodels.com/news/beyondmath-seedWed, 25 Feb 2026 09:00:00 GMTCambridge (UK) startup BeyondMath extended its seed round to $18.5 million, led by Cambridge Innovation Capital with UP.Partners, Insight Partners and InMotion Ventures. The company says its platform rests on a foundation model trained on first-principles physics, so customers need little simulation data of their own; it works with Honeywell on aircraft component design.Aero & CFD https://surrogatemodels.com/news/emmi-neuralmouldhttps://surrogatemodels.com/news/emmi-neuralmouldTue, 10 Feb 2026 09:00:00 GMTEmmi AI presents NeuralMould, a model of the filling stage of injection moulding built with SIMCON, an injection-moulding CAE company. Its model page describes an AB-UPT variant that takes the mould geometry, process parameters, material properties and gate locations, and returns pressure, temperature and shear-rate fields plus pressure and flow-rate curves at the gates. Emmi says it was trained on over a million transient trajectories across thousands of materials, and gives 1 to 2 seconds per prediction against 30 minutes to 1 hour for a traditional simulation, with "consistent 5% relative errors". No paper accompanies it; a demo on the model page asks for a work email.Aero & CFD https://surrogatemodels.com/datasets/shift-pumphttps://surrogatemodels.com/datasets/shift-pumpThu, 01 Jan 2026 09:00:00 GMTOver 5,600 parametric centrifugal pump geometries, impeller and volute, simulated with water in a rotating reference frame at and away from the design flow rate, with surface and volume fields and torque and outlet total pressure for each case. Licence: CC BY-NC 4.0.Aero & CFD https://surrogatemodels.com/datasets/shift-submarinehttps://surrogatemodels.com/datasets/shift-submarineMon, 01 Dec 2025 09:00:00 GMTSimulations of geometric variants of a parameterised Myring-type submarine, six of its more than 20 shape parameters varied, run on the Luminary Cloud platform with adaptive meshing, with surface and volume fields and integrated forces. Licence: Proprietary (Luminary Cloud, all rights reserved).Aero & CFD https://surrogatemodels.com/datasets/shift-ccahttps://surrogatemodels.com/datasets/shift-ccaMon, 01 Dec 2025 09:00:00 GMTSimulations of planform variants of a parameterised Group 5 UAV reference configuration, of the class the card calls Collaborative Combat Aircraft, generated in nTop and run on the Luminary Cloud platform at cruise conditions, with surface and volume fields and integrated forces. Licence: Proprietary (Luminary Cloud, all rights reserved).Aero & CFD https://surrogatemodels.com/market/luminary-series-bhttps://surrogatemodels.com/news/luminary-series-bMon, 15 Sep 2025 09:00:00 GMTLuminary Cloud raised a $72 million Series B led by N47 with Sutter Hill Ventures and NVentures, NVIDIA's venture arm. Its platform pairs simulation that generates synthetic training data with deep-learning models trained on it, which the company says predict how cars, aircraft and electronics perform in near real time.Aero & CFDThermal https://surrogatemodels.com/datasets/shift-winghttps://surrogatemodels.com/datasets/shift-wingTue, 01 Jul 2025 09:00:00 GMTSimulations of thousands of parametric variants of the NASA Common Research Model, with the wing re-lofted from six spanwise sections and the fuselage tube varied in diameter and length, across a range of flow conditions, with surface and volume fields. Licence: CC BY-NC 4.0.Aero & CFD https://surrogatemodels.com/datasets/shift-suvhttps://surrogatemodels.com/datasets/shift-suvThu, 01 May 2025 09:00:00 GMTTransient DDES runs of thousands of variants of the FKFS AeroSUV, morphed with a deformation cage over design parameters informed by Honda, in estate and fastback form at quarter and full scale, with surface and volume fields. Licence: CC BY-NC 4.0.Aero & CFD https://surrogatemodels.com/datasets/drivaernet-pphttps://surrogatemodels.com/datasets/drivaernet-ppSat, 01 Jun 2024 09:00:00 GMT8,150 car designs (fastback, notchback, estateback) with high-fidelity CFD, surface pressure and wall shear stress, volume fields and geometry, released with benchmark code. Licence: CC BY-NC 4.0.Aero & CFD