The right roadmap to defense innovation in space

Artificial intelligence is reshaping the national security landscape, from intelligence analysis to logistics to battlefield autonomy. Yet one of the most consequential forms of AI is still unfamiliar to many outside engineering circles. Physics AI refers to models that learn from data representing the physical world, such as high-fidelity simulations, experiments, and data collected during the operation of engineering systems. These models are built to predict how real-world systems behave.

For example, it could be the stalling point of an aircraft wing, the performance of a jet engine compressor in a dusty environment, or the aerodynamic heating of a hypersonic vehicle moving through the atmosphere.

Although Physics AI sounds like a recent invention, the underlying concept has been studied for nearly 15 years. The first PhD student who worked with me on this topic finished his thesis on this topic in 2013, and we thought of the work as a promising idea whose time was yet to come. Researchers and engineers have long assumed that once data became abundant enough, Physics AI models would be able to replicate and predict physical behavior with relatively high accuracy. For most of the past decade, this vision seemed distant, limited by exceedingly slow simulations and small datasets.

During the last two to three years, however, we have seen an acceleration of Physics AI, an inflection point, that has exceeded my expectations. Three major developments have converged.

First, engineering teams can now generate far larger, higher-fidelity datasets than ever before, mostly based on synthetic data generated by GPU-native, modern simulation tools. Second, the large language model revolution introduced new model architectures and techniques, many of which can be adapted to physical problems. Third, improved data storage, data curation, and data control infrastructure, together with modern APIs/SDKs, have enabled easier model training and sharing across the enterprise. 

These developments produced remarkable results for defense contractors. For example, engineering simulations are currently used to design and manufacture new aerospace products; however, even with the most advanced computing architectures, these processes remain expensive and time-consuming. In contrast, by introducing Physics AI models into engineering workflows, simulations that once required six to eight hours of compute time can now be run in about a second, and without significant compromises in accuracy. Similar improvements are available in other physical disciplines that come together in most modern aerospace systems.

As incredible as that may sound, this shift is just the beginning. Many think that increasing data volumes will be sufficient to realize the full potential of Physics AI in the defense industry. That is not the case. If the defense community wants to achieve the same type of foundational models that are emerging in language AI, more data alone is not enough.

Physical systems follow strict laws and obey specific symmetries and invariants that traditional machine learning does not automatically understand or enforce. Regardless of how much simulation data we generate, there will never be enough real-world data to cover every relevant scenario, and training a model on observed data alone will severely limit its generality.

The next step forward is to embed the physics into the architecture of Physics AI models. We have to make sure that a Physics AI model understands the physics it represents in addition to its uncanny ability to reproduce the data it has seen in training. Doing so will create models that generalize more reliably, behave predictably, and can be applied across a much broader range of aerospace and defense systems.

How can such physics-aware Physics AI models change the way we develop new products in aerospace and defense? When engineers can quickly and accurately test thousands of ideas rather than just a handful, they can look at more alternatives, eliminate the bad ones early, and focus effort on what truly matters. That acceleration translates directly into more capable systems and shorter development cycles. Much as Chinese EV companies have cut new-vehicle development times in half, U.S. aerospace and defense could greatly accelerate the development and deployment of new systems.  This use of cutting-edge AI technology could also help overcome cultural resistance to change in the defense industry, as a younger generation of engineers enters the field and is more familiar with AI tools.

This combination of data-driven learning and physics-informed structure will move the field from impressive demonstrations to transformational capability. It will allow engineers to design systems that would have been impractical to explore through conventional simulation and testing. It will strengthen the competitiveness of defense contractors by expanding the design space, shortening iteration cycles, and improving the performance of mission-critical platforms.

There are legitimate concerns about the race to develop artificial generalized intelligence (AGI) and its possible effects on society at large. Based on my observations, the danger with Physics AI in the defense space is that we will not move fast enough - greatly accelerating innovation while simultaneously increasing safety in design. Introducing Physics AI into engineering development processes will enable the workforce to experience significantly more learning loops, making their work product even more valuable.

The defense community faces a fundamental challenge: as requirements grow more complex, timelines expand, and adversaries seize the opportunity to accelerate their own capabilities and leapfrog the leaders. Physics AI is one of the few levers that can fundamentally change the rate at which the United States conceives, designs, and fields new systems, and always maintains a technological edge. It accelerates development by reducing simulation and design times. It improves quality by allowing engineers to explore the design space more thoroughly. Most importantly, it helps eliminate unworkable ideas quickly, supports risk management, and advances concepts that lead to superior performance. And we’ve barely scratched the surface of its possibilities.

Defense Secretary Pete Hegseth has made Pentagon procurement reform a major priority for accelerating the pace of innovation in military programs. In a recent speech, he called on large defense contractors to “change the focus on speed and volume and divest their own capital to get there.”

Continued progress will require ongoing research, better integration into existing engineering workflows, and the development of models that combine abundant data with embedded physical principles. But the direction is clear. Physics AI will help the defense industry operate at the speed required by the strategic environment. It will enable more rigorous design, faster iteration, and more innovative systems. The nation that learns to use it effectively will hold a significant advantage in future conflicts.