Luminary Introduces First-Ever Agent-Led Experience to Automate the Physics Behind Engineering
News originally published by
GlobeNewswireLuminary Autopilot and Continual Large Physics Models enable faster time-to-value for organizations in their engineering design workflows
SAN MATEO, Calif., Nov. 17, 2026 (GLOBE NEWSWIRE) -- Today at the inaugural Physics AI DevDay, Luminary, the Physics AI company, unveiled two new innovations - Luminary Autopilot, which supports NVIDIA Nemotron, and Continual Large Physics Models (LPMs) which are built with NVIDIA CUDA-X and PhysicsNeMo libraries. These capabilities significantly collapse time to value and expertise requirements and enable highly-efficient Graphics Processing Unit (GPU) usage across the entire Physics AI modeling process. With these developments, Luminary is helping companies accelerate the journey from theory to value by optimizing and simplifying the process of building and launching a Physics AI model to accelerate the design and development process for vehicles, rockets and other physical systems.
Traditionally, building a Physics AI model required multiple steps, including CAD ingestion, geometry cleanup and meshing, and simulations - a process that could take months before a model is even queried. Luminary Autopilot automates and takes an intelligent, rationale-based approach to the process. A single prompt equips agents to carry out work across the platform unattended with well-defined human-in-the-loop approval checkpoints. Specialized agents ingest an engineer’s prompt in the platform or via Slack and then run the associated data generation, training and inference tasks using NVIDIA CUDA-X and PhysicsNeMo libraries. The agent asks for permissions before making changes to data or spending compute, and customers can approve once, always allow a specific command or redirect the agent. Those approvals can be revoked at any time.
“Continuous learning allows engineering AI to keep pace with changing product requirements, turning simulation data into models that grow more capable over time,” said Tim Costa, vice president and general manager of industrial and computational engineering at NVIDIA. “With NVIDIA accelerated computing and PhysicsNeMo, Luminary is helping engineers automate the creation and refinement of physics AI models, reducing the effort needed to apply them to new design challenges.”
Luminary Autopilot leverages the Luminary Data Lake, a new training-ready store that indexes platform job data, third-party simulation data and reference dataset, so that customers can train or fine-tune agents against existing data and public baselines, as well as their own data. The Data Lake is wired directly into Luminary Pipelines, so users can codify the steps their workflows follow, whether they’re launched by a person or an agent.
“Up to this point, setting up a Physics AI workflow required a dozen manual steps spread out over weeks or months and across multiple people. Every handoff introduced the chance for mistakes and took engineering resources away from critical projects,” said Juan Alonso, Chief Technology Officer at Luminary. “Empowering engineers to automate design work simply by describing it to an agent in a single prompt is a significant milestone. Luminary Autopilot and its accompanying capabilities will allow engineers to deliver new or updated designs and, ultimately, new products to market faster while maintaining the accuracy these designs require.”
Customers can also leverage the Data Lake to train or fine tune models without requiring the assistance of a Luminary team member. The public Luminary Software Development Kit (SDK) now includes model training so customers or an agent working on their behalf can point a training job at a dataset in the Data Lake and launch distributed training on platform GPUs. The SDK also enables users to leverage a pre-trained base model and continue training on data from the customer’s domain, delivering domain-specific accuracy at a fraction of the data and compute cost of training from scratch. Customers can choose between full fine-tuning that updates every weight in the model or low-rank adaption (LoRA) that freezes the base model and trains a small add-on network alongside it.
Introducing Continual Large Physics Models
LPMs are becoming increasingly capable and, as a result, organizations are exploring new applications and domains beyond a model’s initial use case. However, these models are limited by their static nature – Once their training is complete, their abilities are fixed and adapting their capabilities requires more data or a new data type and, ultimately, retraining the model.
Luminary is changing this with the introduction of a new, more agile and rapidly-adaptable class of LPMs: Continual Large Physics Models. Building on LPMs’ ability to build highly accurate models that can predict performance characteristics of a design much earlier in the engineering design cycle, Continual LPMs get smarter simply by being used. The models autonomously improve themselves without retraining from scratch and without human initiation, identifying their weaknesses, generating the data needed, and learning from that data without forgetting existing knowledge.
Continual LPMs run in a three-phase loop – active learning, data generation and model training. The process uses uncertainty quantification and out-of-distribution detection to identify knowledge gaps, generate the needed training data, feed information into the Luminary Data Lake, and train a new model version on the data. It retains the model’s existing knowledge and helps it get smarter over time, automatically promotes new models, keeps the outgoing model versioned with full lineage, and gives users the ability to retire or keep any versions as needed without disrupting live inference or other use.
“Customers are demanding that manufacturing organizations deliver new innovations faster and more cost-effectively. Physics AI is helping us accelerate our design time, reduce physical testing and prototyping costs, and reduce field performance uncertainty,” said Suds Menon, Chief Product Officer at Luminary. “The introduction of Luminary Autopilot and Continual Large Physics Models will only amplify our customers’ development engine, illustrating why Luminary is a research and development lab at the forefront of AI innovation.”
Luminary revealed the latest breakthroughs in Physics AI during its Physics AI DevDay virtual event. For more information about Luminary’s most recent to its Physics AI platform, visit https://luminary.ai/resources/whats_new_Luminary_Platform_fall_2026/.
About Luminary
Luminary provides the industry’s only full-stack Physics AI platform, combined with deep industry and deployment expertise, to build Large Physics Models (LPMs) that accurately predict physical behavior in seconds rather than hours or days, enabling customers to reshape their product development processes from design to manufacture and operations. Customers span aerospace and defense, automotive, and industrial manufacturing, including Northrop Grumman, Joby Aviation, and Otto Aerospace. For more information, visit www.luminary.ai.

Media Contact luminary@w2comm.com
Legal Disclaimer:
EIN Presswire provides this news content "as is" without warranty of any kind. We do not accept any responsibility or liability for the accuracy, content, images, videos, licenses, completeness, legality, or reliability of the information contained in this article. If you have any complaints or copyright issues related to this article, kindly contact the author above.