Physics AI as a Strategic Advantage: How Physics-based AI Models Are Reshaping U.S. Defense Engineering
Interview with Juan Alonso Chief Technical Officer & co-founder of Luminary
Ahead of his presentation at CDFAM, and participation in the CD/DC real-time AI-augmented engineering demo for government observers, Juan Alonso, cofounder and CTO of Luminary, discusses the company’s Physics AI models, which replace meshing and solving with a near-instant inference loop.
He covers how the approach generalizes across regimes as different as UAVs, submarines, and supersonic missiles, how engineers validate the results against traditional CFD, and what it takes to connect aerodynamic and structural analysis into a workflow that runs in seconds rather than weeks.
Can you start by telling us a little about your background, your role at Luminary, and what the company is focused on?
I have a long history as an academic and engineer in aerospace. I have been the Chair of the Department of Aeronautics & Astronautics at Stanford University for the past 3 years and served as Director of the NASA Fundamental Aeronautics Program.
My expertise and interest lies in the simulation-based design of all kinds of aerospace systems (aircraft, propulsion systems, launch vehicles, etc.) across the speed regime.
At Luminary, I am the co-founder and CTO where I lead the technical vision of our company to create Physics AI models of all types of systems.
The company as a whole is focused on transforming how physical systems are designed, built, and operated. We want to help engineering teams move faster, explore more possibilities, and make better decisions across the product lifecycle.
We also want to empower engineers, through agentic capabilities, to improve their output by looking at more alternatives, more deeply. Imagine being able to perform multi-objective design optimization of an aircraft by coupling aerodynamics and structural analysis in seconds, as opposed to weeks with traditional simulation methods.
Your models are being applied across domains as different as UAVs, submarines, and supersonic missiles. What properties of the Physics AI approach allow it to generalize across such different flow regimes, and where do you currently see the boundaries of that applicability?
Our model architectures are designed to encode any physical behavior from large amounts of training data which can come from either simulation or experimental data specific to the application and its associated physics. Because the training data typically covers the design space (geometries, flow regimes, boundary conditions), the physics ai model also generalizes in the design space.
More importantly, we are enhancing the data-driven methods often used in AI/ML with true physics constraints that (a) improve accuracy for a given dataset, (b) reduce data requirements to achieve a certain accuracy, and (c) improve generalizability of the models and predictive accuracy in sparse data regions. For these reasons, we feel confident that, in the very near future, Physics AI models are going to be widely applied across various different physics.
Replacing meshing and solving with a near-instant inference loop promises to change the engineering process considerably. How do engineers validate and build confidence in those results against the traditional CFD tools they are more familiar with?
Validation and Verification (V&V) are key in the dataset generation, training, and inference processes: we must make sure that our customers can trust the results produced by the models or, at last, that the models are aware of their deficiencies in limited areas of the design space.
Physics AI models are rigorously tested with unseen datasets during the training process to ensure the model generalizes instead of overfitting on the data. We equip our models with confidence measures to inform the engineers how reliable the predictions are. So along with the physics predictions we provide insights like out of distribution (OOD) detection, uncertainty quantification (UQ).
Lastly, when a model does not provide sufficient accuracy in some areas of the design space, we selectively and adaptively enrich the training dataset with additional simulations and use transfer learning to re-train the models.
Training a surrogate model to accurately capture phenomena like shock interactions and boundary layer effects is a significant undertaking. What does that training dataset look like, is it customer or generic data, and how do you validate accuracy in flow regimes the model has not explicitly seen?
Training data can be synthetic (e.g. simulation) data and/or experimental data. Synthetic data is usually well V&V’d simulation data.
Luminary has proprietary GPU-native solver capabilities to generate the required synthetic data at scale very quickly. Luminary also offers data ingestion capabilities to ingest external simulation data and make it physics ai model training ready.
The models are trained on training data that covers the flow regimes and boundary conditions that the models can be inferred on.
In general, the data is specific to the customer application being pursued. But as we continue to build more generalizable models, with physics constraints, that can predict accurately in regimes that have not been seen before, Luminary is also creating large amounts of generic data (see our SHIFT models, which we release at least once a month) that we use to create foundational models.
Breaking down the silos between aerodynamic and structural analysis is something the industry has been working toward for a long time. What does that integration look like in practice within your workflow, and what does the data handoff between domains actually involve?
Because physics AI models just use the geometry and boundary conditions as an input, they bypass the complicated steps of meshing, solve and convergence that make interactions between different domains (like aerodynamics and structural) cumbersome and human intensive.
For example, an aircraft structural engineer might need a surface pressure distribution based on the flight condition and, instead of waiting for weeks for a CFD expert to perform full fidelity simulation, they can query an aerodynamic physics ai model and have full field outputs in seconds.
In addition, Luminary is building all kinds of robust interpolation techniques for multiple physics (aero-structure, fluid-thermal, fluid-acoustics, etc.) that can be driven agentically. The objective is, together with Physics AI models that bypass meshing, to remove the cumbersome steps for the connection between the physics that have hampered the use of true multi-physics engineering workflows.
What are you hoping to share and take away from the CDFAM audience in DC this year?
I aim to share how the thoughtful combination of Agentic AI and Physics AI is emerging as the most important strategic capability for the Department of Defense and Intelligence Community in the past 3 decades. This will enable the U.S. defense industrial base to move beyond traditional, slow engineering workflows and gain a competitive edge, by reacting more quickly and tailoring their products in a fast-changing world.
By introducing Luminary’s SHIFT models for collaborative combat aircraft, missiles, rocket thrusters, and submarines; I would love to demonstrate that near-real-time, high-fidelity, physics-anchored inference is now essential for next-generation military systems…and it works!
Ultimately, I hope to engage the CDFAM audience on the urgency of adopting these technologies to set the pace for the future of U.S. military engineering, and in better understanding the problems being faced today so that we can help build custom solutions that have sizable impact.
Juan presents at CD/DC in Washington DC on July 15 and 16, alongside others working on physics-based modeling and AI-driven engineering workflows.
The evening program features a fireside conversation between U.S. Representative Chrissy Houlahan and Dr. Will Roper, CEO of Istari Digital and former Assistant Secretary of the Air Force for Acquisition, Technology and Logistics, followed by an Integrated UAS Re-Design and Re-Certification demo. The demonstration shows how five or more fully interoperable, AI-native tools and agents can integrate new sensor and compute capabilities, update an entire UAS design, and re-certify it for manufacturability and safety in under an hour. Register to attend.
