We are entering an era in which artificial intelligence can generate code faster than any engineer, run parametric sweeps in seconds, synthesize regulatory documentation automatically, and couple multiphysics solvers with minimal human input.
I find that genuinely exciting. And I also think most commentary about what this means for engineers in high-consequence fields — nuclear engineering among them — is either catastrophically pessimistic or naively dismissive. The truth is more interesting than either position.
Let me work through it honestly, starting with the uncomfortable question.
The Question Worth Sitting With
When AI systems exceed human capability in knowledge synthesis, code generation, and quantitative output — what value remains uniquely human?
For a nuclear engineer working in safety analysis — particularly one whose career has run through thermal-hydraulic code development, transient analysis, and methods development — this question is not rhetorical. It deserves a careful answer rather than a reflexive reassurance.
The answer is structural, not sentimental.
What AI Can Now Actually Do in Our Field
Honesty first. AI has already crossed thresholds that deserve acknowledgment.
The OECD Nuclear Energy Agency maintains a dedicated Task Force on Artificial Intelligence and Machine Learning for Scientific Computing in Nuclear Engineering,[1] whose mandate spans reactor physics, thermal-hydraulics, fuel performance, and structural mechanics. Physics-Informed Neural Networks (PINNs) — architectures that embed conservation equations directly into the learning process — have been applied to surrogate modeling of accidental scenarios in nuclear power plants.[2] AI-based surrogate models have been developed to accelerate computation for nuclear reactor analysis, including for system-level thermal-hydraulic codes and computational fluid dynamics applications.[3]
On the documentation and licensing side, the pace is accelerating visibly. The U.S. Department of Energy demonstrated that Gordian AI, built on Microsoft Azure, converted the Preliminary Documented Safety Analysis for DOE’s National Reactor Innovation Center Generic High Temperature Gas Reactor into sections equivalent to an NRC license application — producing a 208-page document in one day. Normally, the process takes a team of people between four and six weeks to complete the same task.[4] A study by the National Reactor Innovation Center found that AI has the potential to reduce both document development time and regulatory review cycles by as much as 50 percent, while simultaneously improving accuracy, consistency, and traceability.[4]
At the University of Michigan, the pyMAISE framework — Python-based Michigan Artificial Intelligence Standard Environment — was developed specifically for nuclear engineering applications, with the goal, in the words of its lead researcher, that “pyMAISE is one step to help the NRC create a pipeline for licensable AI."[5]
None of this is science fiction. These tools exist now, in production research environments. The trajectory is clear.
And yet.
The Bottleneck Has Never Been Computation
Nuclear safety analysis is not a throughput problem. It has always been a credibility problem.
The central task is not producing numbers faster. It is producing results that are:
- Physically meaningful within validated regimes
- Defensible under regulatory scrutiny
- Conservative where conservatism is required
- Honest about their own limitations
- Legally and ethically grounded
An AI can compute faster and write documentation faster. It cannot own the consequence of being wrong.
In nuclear engineering, consequences are not abstract. They sit in the regulatory record, in the licensing basis, and ultimately in the public risk calculus. That accountability structure does not dissolve as AI becomes more capable. It concentrates.
Physical Judgment Is Not Pattern Recognition
There is a category of knowledge that experienced nuclear engineers accumulate that does not reduce to pattern matching across training data. It emerges from the specific experience of being wrong — and understanding why.
Safety-focused transient analysis teaches this in a particular way. Consider what it actually means to interrogate a thermal-hydraulic simulation result:
- Recognizing when a validated correlation is being applied outside its experimental basis — not because a flag fires, but because the combination of conditions looks physically wrong
- Identifying when numerical stability is masking physical instability — when smooth output should make you more suspicious, not less
- Knowing when a discretization scheme is distorting phase-change physics in a way that will produce optimistic results under precisely the conditions where conservatism is most needed
AI can interpolate within known regimes. Engineers who have lived through anomalous runs, failed validations, and post-test reconciliations recognize regime transitions — the moments when the physics changes character and the model’s assumptions no longer hold.
This matters especially in beyond-design-basis analysis: the low-frequency, high-consequence scenarios where data is genuinely sparse, where empirical validation is thin, and where physical reasoning dominates statistical inference. These are the conditions under which current AI approaches are structurally weakest. Deep uncertainty is precisely the domain that rewards engineering judgment most heavily.
Recent work on Physics-Informed Neural Networks applied to critical heat flux prediction illustrates both the promise and the constraint: models work well when background physical knowledge constrains the learning, but defining which physical constraints to apply, and recognizing when the model has exceeded its valid range, is itself an engineering act. It requires someone who understands both the physics and the failure modes of the learning system.
That is not a task that can be automated.
The Regulatory Reality Is an Opportunity
The U.S. Nuclear Regulatory Commission is actively working through how to handle AI in licensing. In February 2025, the NRC signaled it was examining “more nuanced requirements” for AI in the nuclear sector, acknowledging that while existing regulations are “fairly flexible enough to adapt to artificial intelligence,” the question of verifying “the explainability, the interpretability [and] all those -ilities of AI” to meet actual regulatory standards “gets a little more nuanced."[6]
The NRC’s regulatory gap analysis found that targeted areas of AI use require establishing specific requirements — “things that need to be shown to prove AI is safe and secure enough."[6]
The international regulatory community has arrived at a parallel finding. The OECD Nuclear Energy Agency’s International RegLab Project concluded that “explainability alone is insufficient for applications with high safety implications,” and that AI systems in this domain must demonstrate quantifiable, auditable justifications to support regulatory confidence.[7]
That determination is consequential. It means regulators are not simply asking whether an AI model can explain itself in natural language. They are asking whether the explanation can be interrogated in the same way a safety methodology can be interrogated — against its validation basis, its conservatism assumptions, its physical interpretation.
Regulators are going to need domain experts to define what “acceptable” means in this framework. The NRC does not have the capacity to develop these standards unilaterally from first principles. The engineers who understand both the physics and the regulatory architecture — who can sit across from a licensing reviewer and defend why a particular surrogate model bound is physically conservative — are going to be at the center of that process.
That is not a marginal role. It is the critical interface.
The Nuclear Renaissance Creates a Different Pressure
There is one more factor that the standard AI-displacement narrative overlooks entirely.
The global nuclear industry is experiencing genuine structural expansion for the first time in decades. The OECD Nuclear Energy Agency’s SMR Dashboard Third Edition identified 127 distinct SMR designs — up from 98 in the previous edition — with 51 designs involved in pre-licensing or licensing processes and 85 active discussions between developers and site owners.[8]
In the United States, the Department of Energy selected TVA and Holtec to receive up to $800 million in federal cost-shared funding — $400 million each — for SMR deployments at the Clinch River Nuclear site in Tennessee and the Palisades Nuclear Generating Station in Michigan, with both projects targeted to deliver new nuclear generation in the early 2030s.[9] China’s Linglong One, the world’s first commercial land-based SMR, is planned to begin commercial operations as early as the first half of 2026, according to Wang Zhenqing of the China National Nuclear Corporation.[10]
Each new reactor design requires a new licensing basis. Each new licensing basis requires safety analysis work — methodology development, code qualification, uncertainty characterization, transient inventory. The analyses performed for existing light water reactors do not transfer automatically to gas-cooled, molten salt, or microreactor concepts. The physics is different. The failure modes are different. The validation experiments have not been run yet.
At the same time, the workforce carrying institutional knowledge from decades of operation is aging. Approximately 60 percent of nuclear professionals in the United States are between the ages of 30 and 54, and a significant retirement wave is approaching in the coming decade.[11] The accumulated knowledge at risk — anomalous run histories, code behavior edge cases, plant-specific peculiarities — is precisely the knowledge that AI systems have no mechanism to absorb.
This is a demand expansion, not a contraction. Engineers who combine deep physical understanding with AI-augmented analysis capability are going to be in short supply in this environment.
The Shift: From Code User to Method Architect
There is a version of professional obsolescence that is real, but it is specific. It occurs not when AI becomes more capable — it occurs when an engineer competes at the wrong layer.
If professional identity is anchored entirely in:
- Running established codes with existing nodalization
- Producing standard safety reports from templates
- Executing predefined analysis methodologies
Then those layers will compress under automation pressure. Not because the work is unimportant, but because it will increasingly be performed faster and more consistently by AI-augmented workflows.
The layer that remains durable — and actually increases in value — is one level up:
- Designing safety analysis methodologies rather than executing them
- Defining the validation matrices that determine when an AI surrogate can be trusted
- Architecting the uncertainty frameworks that govern conservative bounding
- Governing the integration of AI tools into licensed code systems
- Being the accountable human in a licensing proceeding
The future nuclear engineer working in safety-focused roles is not a code user. They are a safety systems architect — the person who defines how AI is allowed to operate, what physical constraints it must honor, and where human judgment must remain in the loop.
A Practical Roadmap
Translating this into concrete action is easier than it sounds, because the directions are well-defined.
The technical foundation for the next decade involves deepening fluency in uncertainty quantification — not just executing established methods, but principled design of uncertainty frameworks for new analysis regimes. The integration of physics-informed learning models into uncertainty propagation is an active research frontier where practitioners with code development background have a genuine contribution to make.
The regulatory engagement is where leverage compounds. Contributing to NRC methodology position papers, participating in licensing working groups for advanced reactor concepts, and developing explainability frameworks for AI-assisted analysis are activities that directly shape the standards everyone else will eventually have to follow. The NRC has explicitly signaled it needs this input from domain experts.[6]
The intellectual positioning is this: stop thinking of the role as “senior analyst” and start thinking of it as the person who sets the rules for what counts as credible analysis. That is a more interesting problem, and a considerably more secure one.
Final Reflection
When AI agents exceed human computational capability — and they will, in most domains — the human engineer does not become obsolete.
The human becomes the framer of questions. The designer of constraints. The guardian of physical truth. The architect of defensible methodology. The bearer of legal and ethical responsibility.
In nuclear engineering, these are not peripheral concerns. They are the core of what the work actually is. The computation has always been in service of a safety argument. The safety argument requires human judgment and human accountability.
As AI grows more capable, that accountability does not diminish. It becomes more visible, more consequential, and — for engineers who choose to occupy that role deliberately — more professionally defining.
The nuclear engineer with deep experience in safety analysis methods, transient physics, and code development is not facing obsolescence.
They are facing an elevation.
The question is whether to step into it.
Sources
OECD Nuclear Energy Agency — Task Force on Artificial Intelligence and Machine Learning for Scientific Computing in Nuclear Engineering
Bakhsh et al., Physics informed neural networks for surrogate modeling of accidental scenarios in nuclear power plants, Nuclear Engineering and Technology (2023) — Korea Science
Mao et al., A review of the application of artificial intelligence to nuclear reactors: Where we are and what’s next, Heliyon (2023) — PMC
U.S. Department of Energy — Department of Energy Unleashes AI to Reduce Reactor Licensing Timelines
Techxplore / University of Michigan — Streamlining AI development for transparent nuclear engineering models (January 2025)
Nextgov/FCW — Nuclear Regulatory Commission to examine more nuanced requirements for AI (February 2025)
OECD Nuclear Energy Agency — International RegLab Project reports on AI use in nuclear power plant operations
World Nuclear News — There are now 127 different SMR designs, finds NEA report (July 2025), citing the NEA SMR Dashboard: Third Edition
U.S. Department of Energy — Energy Department Selects TVA and Holtec to Advance Deployment of U.S. Small Modular Reactors (December 2025)
OilPrice.com — China to Launch First Small Modular Reactor in 2026, citing Reuters / Wang Zhenqing, China National Nuclear Corporation
The Planet Group — Nuclear Jobs Report 2025: Trends, Challenges, and Opportunities in a Transforming Industry