The Agentic Physics AI Platform

Leverage agents to build, deploy, and operationalize Continual Large Physics Models across your engineering programs.

Agents

Your Interface to the Platform

  • Geometry Agent
  • Data Generation Agent
  • Training Agent
  • Visualization Agent

Physics AI

Research & Innovation

  • Model Architectures
  • Physics-Aware Learning
  • Learning & Adaptation
  • Uncertainty & OOD

Physics

Simulation & Experimental Data

  • Geometry
  • Meshing
  • GPU Solvers
  • External Solvers
  • Test Data
  • Visualization

Platform

Core Services

  • Autopilot
  • Data Lake
  • Security
  • Workloads
  • ModelOps
  • SDK & API

Deployment

Same Platform, Every Environment

  • SaaS
  • In-VPC
  • Private Cloud
  • Air-Gapped
  • Hybrid Cloud

Benefits

Physics AI promises to accelerate engineering innovation, but organizations have struggled to realize that promise. Luminary is addressing this with the only full-stack agentic platform for building, deploying, and operationalizing Continual Large Physics Models.


Generate high-fidelity physics data at scale with integrated GPU-accelerated solvers. Integrate existing third-party solvers and physical test measurements, then organize and manage diverse datasets in the Luminary Data Lake to create a unified foundation for model development.


Data generation is built into the platform and unlimited, so teams can run as many simulations as a model needs, from the first training set to every retrain. Geometry, meshing, Luminary's GPU solvers, and the Data Lake run as one pipeline, so results feed training without file transfers between tools.


Build powerful Continual Large Physics Models using advanced architectures and physics-aware learning methods. Capture complex physical behavior, adapt models to new designs and operating conditions, and quantify uncertainty to understand where predictions are reliable and where additional data is needed.


Coordinate AI agents that generate and prepare geometries, mesh and simulate, curate datasets, and train models, autonomously. Accelerate the development and deployment of Physics AI with scalable compute infrastructure, enterprise-grade governance, and flexible deployment options.


Inside the Platform

Every layer runs on one platform, so data, models, and agents share the same storage, compute, and access control.

Agents

Your Interface to the Platform


  • The Geometry Agent de-features imported CAD and fixes gaps, holes, and slivers until the model passes the geometry check and meshes. It extracts internal and external flow volumes for conjugate heat transfer cases. For design studies, it reviews the model's features, asks which ones to parameterize, and generates the variants. Support for mesh-based geometry, such as STL files, with design variants made by morphing the shape, is in development.

  • The Data Generation Agent picks a sampling strategy (Latin hypercube, Poisson disk, factorial, or Sobol) from the stated objectives and composes the pipeline that meshes and simulates each design on Luminary's GPU solvers. It tracks jobs stage by stage, reads logs to diagnose failed runs, and resubmits them. Changes wait for user approval, and results land in the Data Lake.

  • The Training Agent runs the training workflow on the platform: data preparation, configuration, launch, monitoring, and evaluation. It reports decisions and results in plain language, with loss curves and parity plots. Because training is a callable action, active learning and transfer learning loops can start a retrain or fine-tune without a manual handoff.

  • The Visualization Agent turns a plain-language request into a configured 3D scene in Luminary's visualization tool. It selects the field and colormap from the physics context, places cut planes, seeds streamlines, builds Q-criterion isosurfaces, and sets the camera. One message can request a full set of standard views, such as surface pressure and wake structure.

Physics AI

Research & Innovation


  • Teams train Luminary-SMART, Luminary-AB-UPT, Luminary-GeoTransolver, Luminary-DoMINO, and D-JEPA on the platform, or their own architectures. Luminary-SMART reads raw geometry as a surface point cloud, so its predictions do not depend on a training mesh. Luminary-AB-UPT is Luminary's implementation of the Anchored-Branched Universal Physics Transformer published by Emmi AI. D-JEPA, Luminary's own architecture, pairs one geometry encoder with swappable decoders for CFD and structural predictions. New architectures join as the research advances.

  • Luminary-SMART supports several physics disciplines, including conjugate heat transfer and structural analysis, and published SHIFT models predict thermal and structural fields such as module temperatures and Von Mises stress. A modified SMART architecture accepts surface fields as input alongside geometry, so one LPM's output can drive another in a coupled analysis.

  • Measured data enters training as weighted samples alongside simulation data, and a small corrector model learns the remaining gap between prediction and measurement. Pretrained models such as SHIFT fine-tune on a smaller custom dataset for a new design space. OOD and UQ scores show where new simulations add the most information, so new data targets the model's weak regions.

  • OOD detection projects each input geometry and its parameters into an embedding space and returns a confidence score before inference runs, adding negligible time. UQ adds a companion network, about 1% of the LPM's size, that puts an error bar on every surface point and on integrated values such as drag and lift. Held-out simulations calibrate the coverage, such as 90%.

Physics

Simulation & Experimental Data


  • Geometry import accepts continuous CAD and STL files, and Onshape models import directly. Editing operations include union, subtract, and intersect Booleans, translation, rotation, scaling, mirroring, and face imprinting, with an editable history and undo. Automated geometry healing prepares models for meshing.

  • The mesher turns prepared geometry into a computational mesh. Luminary Mesh Adaptation (LMA) then refines the mesh to resolve features such as shocks, boundary layers, and wakes, and repeats this for each point in a design of experiments. Data generation campaigns run without manual mesh iteration between cases.

  • Luminary's GPU-native solvers cover RANS and scale-resolving (DES and LES) CFD from low-speed to supersonic flow, conjugate heat transfer, Ffowcs Williams-Hawkings aeroacoustics for far-field noise, and adjoint sensitivities for shape optimization. Luminary's solver was the fastest in the AutoCFD4 blind comparison, and a 176-million-cell production vehicle case solved in about 20 minutes on GPUs, against a multi-hour CPU workflow. Structural analysis is in development.

  • Results from third-party and in-house solvers load into the Data Lake with Luminary results and train LPMs. Field mapping normalizes each solver's field names to a common physical quantity, so cases from different solvers form one training dataset. Physics outside Luminary solver coverage can still feed a model, because LPM quality depends on the quality of the training data.

  • Wind tunnel results and sensor logs load into the Data Lake next to simulation data. Measurements validate a trained model, enter training as weighted samples when they match the predicted quantities, or train a small corrector model that learns the gap between prediction and measurement. The corrector updates without retraining the main model and covers scalar quantities today.

  • Luminary's visualization tool renders meshes and solutions, computes post-processing filters such as slices, isosurfaces, and streamlines, and runs batch jobs such as animations. Inference through the SDK returns predictions on a plane or in a 3D volume that render in the same tool as simulations. The Visualization Agent drives it from plain-language requests.

Platform

Core Services


  • Autopilot is the top-level agent. It takes a request, sets up its context, and starts subagents, each limited to the context and skills required for its task. Sessions run as durable workflows that survive restarts and pause for engineer approval, and appear on one agents page for monitoring. Code an agent writes runs in an isolated sandbox with scoped credentials.

  • The Data Lake organizes files into Objects, Cases, and Datasets and extracts metadata such as field quantities, units, and statistics at ingest. Engineers filter by fields, parameters, and tags to assemble a Dataset. Freezing a Dataset makes it an immutable record of the data that trained a model, and each training job is checked against that record before it starts.

  • Access control maps each user to the projects, datasets, models, and agents they can use, and Data Lake storage is scoped per account with per-account encryption keys. Luminary holds SOC 2 Type 2 and ISO/IEC 27001:2022 certifications and offers an ITAR-compliant edition. Private cloud releases ship with a deterministic software bill of materials.

  • A Kubernetes-based runtime schedules each meshing, simulation, and training job and allocates GPUs per workload across nodes, so engineers never choose cores, instance types, or GPU counts. Training jobs spread across multiple GPUs with data and model parallelism. In a private cloud, the platform can dispatch compute jobs to an existing SLURM cluster.

  • Each trained model enters a registry with its architecture, weights, hyperparameters, and training metadata, including the dataset version. Retraining creates a new versioned entry and leaves the previous one in place. Models deploy to inference endpoints, and blue-green deployment swaps in a new version, with rollback to the previous one.

  • The Luminary Python SDK scripts geometry, meshing, simulation, pipelines, model training, inference, and the Data Lake, so a team can run the full workflow from code. An Agent API is in development that will let external systems and local agents start, manage, and converse with Luminary's hosted agents through the SDK.

Deployment

Same Platform, Every Environment


  • Luminary runs the platform on Google Cloud and manages the infrastructure, so teams start with only a browser. An ITAR edition runs on Google Assured Workloads for export-controlled programs. The SaaS platform completed a SOC 2 Type 2 audit covering security, availability, and confidentiality.

  • The platform self-installs in your AWS account on Amazon EKS or Red Hat OpenShift Service on AWS, with S3 storage and managed Postgres. It runs the same codebase as SaaS, so the capabilities match. Data, model weights, and runtime execution stay under your control.

  • Luminary Private Cloud runs on Red Hat OpenShift or Kubernetes, with a customer-provided Postgres database and S3-compatible object store. It delivers the same simulation, model training, and inference capabilities as the hosted service. Compute-heavy jobs can run on an existing SLURM cluster, so the platform uses GPU and CPU nodes already in place.

  • An air-gapped deployment is Luminary Private Cloud with no external network access. Luminary ships container images as archives that are staged inside the perimeter, and the platform's agents run on a supported LLM that the customer hosts locally. Luminary recommends Red Hat OpenShift for these sites because of Red Hat's record of patching vulnerabilities.

  • An organization can run more than one deployment at once, for example on premises for secure programs and in the cloud for extra capacity. Every deployment runs the same codebase, so the same workloads run in each with no code changes. Policy-driven bursting and data migration between deployments are in development.

Works with the tools your team already uses.

Design and collaboration

Onshape logonTop logoBlender logoSlack logoMicrosoft Teams logo

Models and infrastructure

NVIDIA PhysicsNeMo logoGoogle Cloud logoAWS logoRed Hat OpenShift logoKubernetes logo

CAE solvers

Ansys logoSiemens Simcenter STAR-CCM+ logoOpenFOAM logoSIMULIA Abaqus and CST Studio Suite logoHexagon MSC Nastran logo

Models that improve as engineers use them.

A closed-loop to build Continual Large Physics Models (CLPMs) that autonomously learn from both new data and their own predictions to improve in accuracy and generalizability over time.

  1. Generate High-Fidelity Physics Data

    Sample the design space, generate geometries, and prepare and mesh them for simulation. Run Luminary's GPU-native solvers or the solvers your team already trusts, add test data, and store it all in the data lake with full lineage.

    Geometry & Mesh

    design-space sampling

    ↓

    Simulation & Test Data

    solver + test runs

    ↓

    Data Lake

    experimental + synthetic

  2. Turn Physics Data Into a Predictive Model

    Curate a training set from the data lake and train a Large Physics Model on multi-GPU infrastructure. The model trains on full 3D simulation fields, grounded on experimental measurements and operational data.

    ↓

    Training

    Multi-GPU training

    ↓

    Continual Large Physics Model

  3. Agents Run Each Step

    The Autopilot agent hands each step, from geometry preparation to training, to a subagent which handles the manual, time-consuming work. Engineers have full observability into every action and control over what runs.

    Autopilot assigns each step to an agent

    • Geometry Agent→ Geometry & Mesh
    • Data Generation Agent→ Simulation & Test Data
    • Visualization Agent→ Data Lake
    • Training Agent→ Training
  4. Predictions in Seconds

    The model takes geometry, boundary conditions, and operating conditions as inputs, with no meshing or preprocessing. It returns the 3D field, integrated quantities such as lift and drag, and quantified uncertainty.

    ↓

    Inference

    3D fields + quantified uncertainty

  5. Fill Gaps in the Training Data

    When an input falls outside the training data or a prediction carries high uncertainty, the model generates new ground truth around that input, from simulation or test, and retrains the model. The model keeps improving as engineers use it.

    ↓

    Generate targeted ground truth

    ↑

From days to seconds at a global automotive OEM.

A global automotive OEM gave designers aerodynamic feedback directly within their design tool via a CLPM. Iterations that took three to four days per design change now take seconds.

1–3sAero feedback, down from 3–4 days
~1%Mean error vs. CFD, held-out set
<5Minutes for CFD on flagged designs

Predictions engineers can defend.

Each prediction comes with what an engineer needs to decide whether to act on it.

Out of distribution

Out-of-Distribution Detection

The platform flags designs that fall outside the training data, so engineers know when a prediction is an extrapolation before they rely on it.

Mach numberPredictionLow uncertaintyHigh uncertainty

Uncertainty Quantification

Each prediction includes error bars across the surface and volume, so engineers can judge whether it is accurate enough for the decision at hand.

Simulation onlyTest measurementsGrounded model

Model Grounding

Models can be grounded on your physical test measurements, with measurement error taken into account, for more accurate predictions.

PredictionModel versionTraining setSource dataValidationMetadata

Secure Traceability

The platform records the source data, model version, and validation behind every prediction, so engineers can defend the decisions built on it.

Our engineers start it. Your team runs it.


  1. With Luminary engineers

    Stand Up the Capability

    Luminary's forward-deployed engineers stand up the first capability with your team. They build an initial Large Physics Model on your data and deploy the platform where your program data is allowed to run.


  2. Your team, on its own

    Run It Self-Service

    Once the first model and the platform are in place, your engineers can leverage the platform. They build new models, add data, and retrain on the platform themselves, without requiring the involvement of a Luminary engineer.

FAQ


No. Results from the external solvers your team already trusts go into the same data lake and training sets as results from Luminary's GPU-native solvers. Every capability sits behind one API, so you can bring your own mesher or CAD cleanup and pick up the workflow from there.


No. The platform includes unlimited data generation, so teams can run as many simulations as a model needs, for the first training set and for every retrain. In Luminary's cloud, the GPU compute for those simulations is billed at usage. In your own environment, they run on your hardware.


Analysts build and validate a model, and downstream teams use it without running simulations themselves. Designers check the aerodynamic effect of a shape change before handing it off, structures teams pull predicted loads into their loads databases, and controls engineers evaluate points across the operating envelope.


Luminary holds ISO/IEC 27001:2022 certification and a SOC 2 Type 2 report for its SaaS platform, is CMMC certified, and complies with NIST 800-171. Programs under ITAR run in a separate ITAR deployment on Google Assured Workloads, and programs that need full isolation can deploy the platform in their own environment. Inside the platform, access control covers projects, datasets, models, and agents, and data owners grant and revoke access to their data. Each subagent the Autopilot starts is limited to the context and skills its task needs. The SOC 2 report, security policies, and penetration test results are available on request. Visit the Luminary Trust Center


Luminary hosts the platform as SaaS on Google Cloud, with an ITAR variant on Google Assured Workloads. You can also install it in your own cloud account, with native integrations for AWS, or on your own Kubernetes cluster in a private cloud or on premises. Private cloud deployments are certified on Red Hat OpenShift, and upstream Kubernetes is also supported. The platform installs into its own namespace in an existing cluster and signs users in through your identity provider, such as Okta.


Yes. The same platform runs in every deployment, so an organization can run on premises for secure programs and in the cloud for extra capacity without changing its workloads. Policy-driven bursting, which moves compute jobs and data lake data between deployments automatically, is in development.


The agents need an LLM, and an air-gapped environment has no connection to a hosted one. You host one of the LLMs Luminary has tested on the platform inside the environment, alongside the platform's database and object store.