DimensionLab / Engineering & business briefing
Physics AI.
Inside your engineering workflow.
Frontier connects simulation data, AI surrogate development and engineering applications. It gives simulation teams and AI researchers a shared place to prepare datasets, run experiments, inspect fields and put trained models to work.
In the DrivAerML example, Frontier evaluates NVIDIA’s pretrained GeoTransolver on 48 held-out car designs and retains 20 cases for field inspection. Mean pressure error is 2.81% relative L1; mean drag error is 0.84%. Inspect the evaluated designs.
Who it is built for
A shared workflow across engineering and AI.
CAE methods & simulation teams
Turn simulation archives into reusable datasets. Preserve the geometry, conditions and preparation choices that make a result meaningful.
Research & Physics AI teams
Bring your own models and training code. Compare experiments on a consistent data split and connect the metrics to physical fields.
Validation & technical leads
Inspect predictions, local errors and engineering quantities. Keep evaluation artifacts available for a model review.
HPC & platform teams
Fit execution to existing compute, storage and operating practices. Define the path from workstation development to scheduled or remote jobs.
From a simulation campaign to a reusable model
One experiment, with its context attached.
Connect and inspect
Import case data, preview geometry and fields, and choose the samples and physical quantities that matter to the study.
Prepare and develop
Record transformations and splits in a recipe. Train a prepared architecture or integrate an existing Python implementation.
Compare and review
Examine held-out predictions on geometry. Compare run metrics, error distributions, uncertainty and the relevant physical checks.
Integrate and reuse
Expose inference to an engineering application. Carry the preparation, configuration and evaluation method into the next dataset or design family.
Connections to the software you already use
Keep the solver.
Connect the learning workflow.
Frontier sits between simulation output and model use. Integration can begin with exported files and Python code, then extend to connected solver sessions, scheduled jobs and application endpoints.
Case parameters
Measured data
Inspect · evaluate
Package · serve
Internal applications
Model review
Simulation & CAE
OpenFOAM · Ansys Fluent · VTK · CGNS · EnSight
Read solver exports into datasets with geometry, fields, units and case metadata. Use vendor-side exporters or a live-session connector where the installation supports it.
Research code
Python · PyTorch · NVIDIA PhysicsNeMo
Instrument an existing training loop through the SDK, or use a registered model in a Frontier pipeline. Retain configurations, metrics, checkpoints and field artifacts.
Engineering applications
HTTP inference · Python clients · ONNX Runtime
Call a trained model from an application or script. Package supported models for a target runtime and check exported predictions against the original checkpoint.
Cluster & cloud compute
Local workers · cloud GPUs · SLURM
Select a job target, stage the inputs, and bring logs and artifacts back to the experiment. Configure the cluster account, partition, runtime and shared working directory.
Custom tooling
Connectors · transforms · model and evaluation plugins
Extend the pipeline around internal file formats, preparation methods and engineering quantities. Keep domain-specific methods in your own Python packages.
Work locally / Execute where it fits
Separate the workspace from the compute.
Engineers can develop and inspect results on their workstation while a selected worker executes demanding jobs. A deployment design establishes where project metadata, simulation data, checkpoints and services live.
Workstation + GPU workers
Keep the Frontier server local and attach remote compute for selected work. Use a dedicated GPU environment for training or a serverless inference endpoint for prediction workloads.
Customer-hosted server
Place the server and workers in the chosen environment. Define access to shared data, artifact storage, backups and service endpoints with the platform team.
Existing HPC
Submit through SLURM using a site profile: login-node access, account, partition, runtime and shared filesystem. Preserve the scheduler as the job execution authority.
On-premises & isolated networks
Plan installation bundles, approved dependencies, data transfers and updates for the site. Scope enterprise identity, project permissions and operational acceptance with IT.
Where a first workflow can lead
Start with one use case.
Build a reusable engineering capability.
Across simulation campaigns
Reusable methods
Apply an established preparation and evaluation recipe to another dataset. Extend it for a new design family, solver output or model architecture.
Into design exploration
Application integration
Connect inference to parameter studies and optimization scripts. Use surrogate predictions to screen candidates and route selected designs back to the simulation process for verification.
Into operational tools
Project-specific development
Develop integrations for a digital twin, monitoring workflow or controls study. Define the inputs, update rate, deployment runtime and validation criteria for that application.
Across the organization
Enterprise integration planning
Link the workflow to internal catalogs and PLM identifiers, define model review responsibilities, and build custom connectors around the team’s systems of record.
A practical evaluation
Measure value on your own engineering case.
Choose a representative dataset and a decision the surrogate should support. Agree the accuracy criteria and operating range, then compare the complete workflow with the current process.
Engineering effort
Time spent preparing data, configuring experiments, reviewing results and handing the work to another engineer.
Experiment turnaround
Elapsed time from simulation exports to an evaluated model, including data movement, queues and execution.
Prediction quality
Held-out field and engineering-quantity errors, relevant physical checks, uncertainty and performance across the intended operating range.
Operational fit
Repeatability on the chosen infrastructure, inference integration, compute and storage usage, and the path to the next campaign.