Luminary Autopilot Brings Prompt-Driven Physics AI Workflows
Simulation software developer Luminary, formerly Luminary Cloud, has introduced Luminary Autopilot, a physics AI tool that automates the creation of models that predict simulation results. It launched alongside a second product, Continual Large Physics Models (LPMs). Engineering.com reported the launch on November 18, 2026.
- Prompt-driven setup: Luminary says Autopilot builds a custom physics AI workflow from a single prompt. AI agents then handle data generation, training and inference.
- Nvidia stack: The agents run on Nvidia's CUDA-X and PhysicsNeMo libraries, and Autopilot supports Nvidia's Nemotron models.
- Luminary Data Lake: A new store indexes platform job data, third-party simulation data and reference datasets. Customers can train or fine-tune against their own data and public baselines.
- Availability: Both products are slated for Luminary's Fall 2026 release. Pricing has not been announced.
How Continual LPMs differ
An LPM learns from thousands of physics simulations and then estimates how a new design will behave without solving the governing equations again. Luminary CEO Pete Schlampp describes these models as the core of physics AI. Standard LPMs stay fixed after training and must be retrained to take in new data.
Luminary says its Continual LPMs improve through use. According to the company, they spot their own weak areas, generate the data they need and learn from it without losing earlier knowledge or needing a person to start the process. Luminary's published examples so far come from aircraft wings, EV batteries and vehicle crash safety rather than buildings.
Our take: what AEC teams should check before trying Luminary Autopilot physics AI
Where it could fit
AEC firms run repeated CFD and thermal studies for wind loads, pedestrian comfort, natural ventilation and data center cooling. A trained surrogate could make early option studies much faster, letting designers compare dozens of massing or layout options instead of a handful. Luminary has not shown AEC use cases, though, so firms would be building their own training sets. Teams weighing options should also look at SimScale, whose engineering AI blueprint for scaling AI agents is likewise built around Nvidia tools.
Open questions
Self-improving models raise a traceability issue. If a Continual LPM changes between concept and design development, the same input could produce different results, which is awkward for engineers who sign off on performance. Ask whether model versions can be frozen per project and how accuracy against full solver runs is reported.
The Data Lake needs scrutiny too: who owns indexed project data, and whether it informs models used by other customers. Pricing, geometry import from BIM tools and the degree of dependence on Nvidia hardware are also not yet public.
A sensible first step
Once the Fall 2026 release ships, a pilot on a completed project with validated CFD results would show quickly whether the predictions hold up for building-scale problems.
Recent Articles
Learn about the latest architecture software, engineering automation tools, & construction technologies
