Building the science of Physics AI

SHIFT-Truck and transfer learning across vehicle classes

A pickup-truck aerodynamics model, about 3,000 times faster than CFD, with a median drag error of 3.8 counts. Starting SHIFT-Truck from SHIFT-SUV improved surface-field predictions at every dataset size tested.

Physics AI–enabled aerodynamic databases (AeroDB)

A parent model trained on 1,000 simulations adapts to a new geometry with about 80 targeted solves, across Mach 0.1 to 8 with prediction error below 3%.

Out-of-distribution detection and uncertainty quantification for Large Physics Models

A Large Physics Model maps geometry and operating conditions to the resulting physical fields. Luminary-OOD flags when an input falls outside the range the model was trained on, so engineers know when to go back to full simulation. It does this without the roughly 50x inference cost of Monte Carlo dropout.

Luminary-SMART: a mesh-independent architecture

A model architecture with multi-scale geometric grounding that works directly from geometry, without a simulation mesh. It trains up to 3.2x faster than Luminary's GeoTransolver implementation and was the most accurate of the architectures Luminary tested, including DoMINO and GeoTransolver.

Visualization and Analysis Agent

Engineers explore simulation results by asking questions in plain language.

Grounding Physics AI models in physical measurements

Grounding means correcting a model's systematic error using test data. Calibrated against a handful of wind-tunnel runs, predictions reach R² values between 0.94 and 0.97.

SHIFT-Battery: thermal design of battery cold plates

The first SHIFT model for conjugate heat transfer. The per-design cycle drops from about 80 minutes to about 30 seconds.

US Patent 12,597,205 B2 granted

"Hybrid GPU-CPU Approach for Mesh Generation and Adaptive Mesh Refinement."

SHIFT-Crash: full-vehicle crash prediction

A model for structural crash dynamics with under 3% root-mean-square error. Inference takes about 50 seconds on a single NVIDIA T4 GPU. The dataset is published on Hugging Face.

SHIFT-Missile: supersonic aerodynamics

Trained on about 6,000 high-fidelity simulations from Mach 1.5 to 3.0.

Luminary Pipelines (Early Access)

One call runs hundreds to thousands of design variants through geometry, meshing, simulation, training, and inference.

NVIDIA GeoTransolver adopted across SHIFT models (NVIDIA GTC)

GeoTransolver is a transformer architecture with geometry-aware attention. Retraining the SHIFT models on it improves prediction accuracy by up to 10x.

Luminary Private Cloud

The Physics AI platform runs fully air-gapped, inside the customer's own data center, with no external connection.

SHIFT-CCA, SHIFT-Submarine and SHIFT-Pump (AIAA SciTech)

Three models for air, sea and turbomachinery, delivered through a secure, U.S.-based Physics AI Factory.

• – SHIFT-CCA: about 4,000 geometry variants; lift R² of 0.990.
• – SHIFT-Submarine: sub-percent drag error.
• – SHIFT-Pump: more than 5,500 simulations; R² of 0.996.

Combining and comparing adjoint methods and Physics AI for aerodynamic shape optimization on GPUs (AIAA SciTech)

A peer-reviewed side-by-side study of gradient-based adjoint optimization and Physics AI models, run on the same GPU-native platform.

US Patent 12,511,173 B2 granted

"Cloud-Based Framework for Analysis Using Accelerators."

Public interactive Physics AI demo

Anyone can run predictions from a model trained on the SHIFT-Wing dataset.

Physics AI model for spacecraft structures, with Northrop Grumman (NVIDIA GTC DC)

Luminary's first model beyond fluid dynamics. It returns structural results more than 3,000 times faster than traditional simulation, with a median error of 5.2%.

Onshape by PTC integration

Parametric CAD, simulation and Physics AI model development in one loop.

SHIFT-Wing: open dataset and model for transonic wing design

Developed with Otto Aviation and NVIDIA, on a NASA Common Research Model–based design space. Drag predictions are within 2% for in-distribution designs. By December 2025 the dataset had grown past 27,000 simulations, the largest open dataset for transonic wing design.

Luminary MCP server and AI Notebook

AI coding agents in VS Code, Cursor and Claude connect directly to the platform to set up simulations and generate data. A hosted notebook pairs an AI assistant with the Luminary SDK.

The Physics AI Factory

One platform for generating simulation data, curating it, training models and serving predictions.

SHIFT-SUV: open dataset and model for SUV aerodynamics

The first model in the SHIFT family, developed with Honda and NVIDIA. It was trained on more than 1,000 high-fidelity simulations and predicts drag with about 3.45% average error. The dataset is published on Hugging Face.

Blender integration for Physics AI dataset generation

Designers launch simulations from Blender and generate training datasets for Physics AI models from geometry built there.

Up to 50x faster Physics AI training, contributed to NVIDIA PhysicsNeMo

An open-source contribution cuts training epochs by up to 50x and speeds data loading 8x. A model that took nearly five days to train now trains in about two hours.

Physics AI design optimization with nTop and NVIDIA PhysicsNeMo (NVIDIA GTC)

Implicit-geometry design in nTop is linked to Luminary simulation data and NVIDIA PhysicsNeMo models, so model predictions drive the shape-optimization loop.

Exascale visualization infrastructure (IJHPCA)

Luminary researchers co-author the DOE Exascale Computing Project's ALPINE paper on in-situ and post-hoc visualization, published in the International Journal of High Performance Computing Applications.

GPU-based adaptive solution technology for the 5th AIAA High Lift Prediction Workshop (AIAA SciTech)

A peer-reviewed paper presenting Luminary's GPU-native, mesh-adaptive results on the high-lift benchmark, with a companion V&V report.

Keynote at the GPPS Forum, ETH Zurich

"Making NASA's Vision 2030 CFD a Reality 6 Years Early," delivered by CTO Juan J. Alonso.

A virtual wind tunnel running at interactive speed, built with NVIDIA Omniverse (SC24)

Shown at Supercomputing 2024 as part of NVIDIA's Omniverse Blueprint for interactive physics digital twins.

Adjoint sensitivity analysis on GPUs

The fully differentiated GPU solver computes design sensitivities for gradient-based shape optimization.

AutoCFD4: a 157-million-cell scale-resolving simulation in 93 minutes

At the 4th Automotive CFD Prediction Workshop, Luminary ran a full transient DrivAer simulation in 93 minutes on 48 NVIDIA A100 GPUs, and in 44 minutes on 64 H100 GPUs.

NASA Common Research Model validation

Lift and drag predictions on the international standard transonic benchmark are compared against participants in the 6th AIAA Drag Prediction Workshop.

Francis-99 hydraulic turbine validation

The solver is validated against experimental measurements of discharge, head, torque and efficiency across the turbine's full operating envelope.

Deep learning for interactive scientific visualization (IEEE TVCG)

"Interactive Isosurface Visualization in Memory Constrained Environments Using Deep Learning and Speculative Raycasting," in IEEE Transactions on Visualization and Computer Graphics. It is extended from a Best Paper at LDAV 2023.

Lumi Mesh Adaptation and conjugate heat transfer

Automatic, physics-informed mesh adaptation removes manual meshing from the simulation loop. That is the step that otherwise limits how many high-fidelity runs a team can generate. The same release extends the platform to coupled fluid-thermal analysis.

ITAR-compliant edition

An isolated, export-controlled environment for U.S. aerospace and defense programs.

GPU-native simulation platform launches publicly

Luminary comes out of stealth with a cloud platform whose solver was built from scratch for GPUs. It runs simulations about 100 times faster than legacy CPU-based tools.

Luminary researchers chair ISAV at SC23

Luminary researchers serve as General and Program co-chairs of ISAV, the peer-reviewed workshop on in-situ analysis and visualization at the Supercomputing conference. They chair it again at SC24, where CTO Juan J. Alonso also gave a plenary talk.

First public validation of the GPU-native solver

Luminary's solver results for the DrivAer vehicle at the 3rd Automotive CFD Prediction Workshop (AutoCFD3) are compared against workshop participants and wind-tunnel data.

Two foundational patent applications filed

The applications, first filed in June 2022, cover a cloud framework for engineering analysis on GPU accelerators, and a hybrid GPU-CPU method for mesh generation and adaptive mesh refinement. Both have since been granted (see December 2025 and April 2026).

Leadership of the SU2 Foundation and the annual SU2 Conference

Luminary researchers serve on the board of the SU2 Foundation, a non-profit behind one of the most widely used open-source CFD and design-optimization codes. They also organize and peer-review its international conference, now in its seventh year.