Applied Research Associates, Inc. (ARA) will spotlight its defense innovation expertise at the 2026 Space and Missile Defense Symposium, taking place Aug. 11–13 in Huntsville, Alabama.
ARA’s advanced nuclear effects modeling capabilities will be among the key technologies featured at the event. To provide insight into this evolving field, Jay Kerwin, an ARA senior scientist specializing in high- and low-altitude nuclear environments, answered questions about the growing role of artificial intelligence (AI) and machine learning (ML) in nuclear environment modeling.
How are AI/ML being applied to nuclear environment modeling and how can these next-generation capabilities improve on traditional methods?
Short answer: simulation data organization into physically meaningful predictions within operationally relevant timescales for wargaming analysis.
Longer answer: We have to acknowledge the history of computer technology and computing power here. In decades past, approximations to fewer dimensions or simpler physical models were essential to allow for direct numerical simulation of problems at hand. With the advances in computing power we now find ourselves in a data driven bottleneck instead of a physics simulation capability bottleneck, though with modern cutting edge problems we can still require days to weeks to months of compute time. The historical approach to enable fast analysis given the bottlenecks created by large amounts of data and long compute times was to create databases that sufficiently span the input parameter space that can be queried at run time with interpolation to provide required information at requested location in input parameter space without running an entire simulation at time of request. Again, this approach worked when input parameter spaces and density of required calculations within the parameter spaces were relatively small. With highly multi-dimensional input parameter spaces we now find ourselves making great progress in training machine learning (ML) models to represent 3D, 4D, 5D, etc… output data streams in a non-interpolative fashion. Our ML models retain physical realism in output predictions in regions of the parameter space where interpolation will simply not work. Some keys to success we’ve found in our current approaches are clever choices for coordinate spaces, identifying invariant spaces, and leveraging symmetries.
Another large area of value and capability afforded by AI is refactoring and cleaning up legacy code bases. Tools that have been developed over many decades bear the marks of multiple standards and are not always self-consistent. We have a responsibility to push codebases towards readability, maintainability, and portability. AI tools have already proved helpful in this regard.
AI and ML are becoming more common across many industries. How are they being used in nuclear environment modeling and why is this an important area of investment now?
I think we have a responsibility as stewards for government owned modeling and simulation tools to embrace both efficiency and security. The efficiency piece is wildly apparent when using agentic tools for tasks like legacy code mapping and new code development. Efficiency here clearly translates to increased value for the government customer. We can develop far more capability in far less time and then focus our efforts on the verification and validation process. While the upside is clear, we must be vigilant about ensuring security as well. Robust agentic capabilities don’t necessarily need to be stood up by the government, though they do need to be vetted and approved at the government level in order for the entire defense industrial base to engage with these tools to realize their benefits in a way that minimizes security and information vulnerabilities.
How can AI/ML improve nuclear modeling capabilities and support better decision making for government and mission stakeholders?
To cut right to it, I don’t think it will ever be possible to ethically put escalation or real-time mission decision making in the hands of an AI tool. What AI/ML can provide is an incredibly high-throughput, multi-dimensional analysis of data and information to a level that humans simply can’t. It seems evident to me the value in AI/ML processing for decision making, either planning or in real time, will be realized in high level lists of options with detailed likely outcomes to be processed by a human-in-the-loop.
What are some of the key challenges facing the use of AI/ML in nuclear environment modeling?
It is not acceptable, as scientists who work in modeling and simulation, to use a “black box” tool where we can’t analyze and critique the logical flow that leads to a conclusion. We can use a “black box” tool, but the onus then lies upon us to rigorously verify the conclusions are sound. This leads to what I’ll call static ML models. For static ML models that are trained, verified, then utilized, we have a low risk for catastrophic misrepresentation by the model. For real-time train and deploy type models, there is always a risk, albeit small, of catastrophic model misrepresentation of reality.
How is ARA addressing these challenges and advancing the use of AI/ML in nuclear environment modeling now and into the future?
At ARA we believe in the power of AI/ML tools and we are building our team’s mastery and command of these tools to expand the space where “black box” approach meets custom solutions yielding both insight and speed. This growing expertise will enable us to continue delivering value to customers and stay up to date with current technology.
Visit https://www.ara.com/smd/ to learn more and connect with us at SMD.
JOIN US
Booth 831
CONTACT
SMD-2026@ara.com
