Nuclear engineers

Three parts of your work where AI is already doing real lifting, and what stays yours.

• Slide rule1942–1960

  Slide rule, hand calculation, and experimental reactor data (Manhattan Project era)

  The physicists and engineers of the Manhattan Project designed the first reactors using slide rules, hand calculation, and experimental data from critical assemblies. Fermi's design of CP-1 relied on careful measurements of neutron multiplication from subcritical graphite-uranium piles, extrapolating by hand to the critical configuration. Shielding calculations were done analytically using the moments method and limited experimental data. Thermal-hydraulics was essentially empirical: reactor cooling was understood by analogy with boiler engineering, and critical heat flux data came from experiments on heated surfaces, not codes.

• Mainframe deterministic codes1957–1985

  Mainframe deterministic codes: ENDF nuclear data, LASL codes, early IBM mainframe calculations

  The first nuclear engineering software era arrived with mainframe computers in the late 1950s and 1960s. The Evaluated Nuclear Data File (ENDF) project, begun in 1964 at Brookhaven National Laboratory, created the first standardized, computer-readable nuclear cross-section libraries that all subsequent neutronics codes would draw on. Los Alamos wrote the first versions of MCNP (Monte Carlo N-Particle) in Fortran in the late 1950s and early 1960s; the IBM 7094 and later the CDC 6600 made it practical for reactor shielding design. The ORIGEN code (Oak Ridge Isotope GENeration) appeared in 1973, enabling burnup and decay-heat calculations. These tools allowed nuclear engineers to move from purely experimental methods toward simulation-based design, though run times measured in hours on mainframes constrained how many parametric variations could be explored.

  Effect on the work

  Mainframe codes dramatically increased the analytical productivity of nuclear engineers compared to hand calculation but required specialized programming skills and computer time allocation that remained bottlenecks through the late 1970s.

• Post-TMI regulatory computing1979–2000

  Post-TMI regulatory computing: NRC-approved codes, RELAP5, CONTAIN, PRA frameworks

  Three Mile Island transformed what nuclear engineers are required to do. The NRC's response to TMI included mandating probabilistic risk assessment (PRA) for all operating nuclear plants (NUREG-1150, 1987-1990), requiring best-estimate thermal-hydraulics analysis using NRC-approved system codes (RELAP5, TRACE, TRAC), and building out the Quality Assurance framework (10 CFR 50 Appendix B) that governs all safety-class engineering work. RELAP5 (Reactor Excursion and Leak Analysis Program, developed at Idaho National Laboratory) became the industry standard for loss-of-coolant accident and transient analysis; its validation and use for licensing-basis calculations became a core nuclear engineer skill. The post-TMI regulatory environment also created the nuclear licensing documentation workload that now consumes a substantial fraction of a nuclear engineer's time.

  Effect on the work

  Plants that might otherwise have operated with smaller engineering staffs needed substantially larger ones to satisfy the post-TMI NRC regulatory requirements.

• Workstation-era Monte Carlo1995–2015

  Workstation-era Monte Carlo and coupled codes: MCNP5, Serpent, SCALE 6, digital plant systems

  The transition from mainframe to engineering workstation computing in the mid-1990s transformed Monte Carlo neutronics from a once-a-month mainframe run into a routine daily tool. MCNP5 (LANL, 2003) ran on Linux workstations and could complete a full-core shielding calculation in hours rather than days. Serpent (VTT, 2004) was optimized for group-constant generation for deterministic reactor simulators, enabling the detailed lattice-physics calculations underpinning reload design. SCALE 6 (ORNL, 2009) integrated criticality safety, shielding, and spent-fuel analysis in a unified framework. Digital plant systems (Westinghouse BEACON, GE GENESISTM) enabled real-time core monitoring at operating plants, shifting a portion of nuclear engineer work from after-the-fact analysis to continuous plant health surveillance.

  Effect on the work

  Workstation computing sharply reduced the marginal cost of each parametric study, allowing individual nuclear engineers to conduct analyses that previously required teams. This contributed to the productivity-led workforce reduction visible in the 2000s OEWS data, where total employment remained in the range despite a fleet of more than a hundred operating reactors.

• High-performance computing2015–2023

  High-performance computing and digital twins: cloud HPC, Bentley iTwin, COMOS plant systems

  Cloud high-performance computing and digital twin platforms brought two qualitative changes to nuclear engineering practice in the 2010s. Cloud HPC (AWS, Azure HPC instances) made large-scale Monte Carlo calculations accessible without national laboratory time allocation: a calculation that required a supercomputer in 2000 could run overnight on a cloud cluster by 2015. Bentley iTwin and Siemens COMOS brought plant engineering into a continuous-model environment: the nuclear plant's physical state, design documentation, and sensor streams were unified in a digital representation that engineers could query in real time. For SMR development organizations (which began serious design activity in this era), simulation-first workflows became the norm: reactor geometries were created as CAD models, fed directly into neutronics codes via DAGMC interfaces, and iterated rapidly without building physical mockups.

• GPU-accelerated Monte Carlo2023–present

  GPU-accelerated Monte Carlo and AI plant monitoring: OpenMC GPU, Westinghouse Lumen, INL RAVEN

  The current tool era is defined by two parallel developments: GPU acceleration of Monte Carlo neutronics codes, and AI-powered plant monitoring. OpenMC (MIT, GPU-accelerated release 2024) reduces full-core Monte Carlo runtimes by 10-100x compared to CPU-only calculation, making what was previously a multi-day supercomputer job a same-day workstation task. MCNP6.3 (LANL, 2023) added improved parallel MPI-plus-OpenMP performance and enhanced variance reduction for deep-penetration shielding. Concurrently, Westinghouse Lumen and GE Vernova SmartSignal deployed deep-learning plant health monitoring across the operating fleet, processing hundreds of sensor streams in real time and surfacing early degradation signatures weeks before conventional alarms would trigger. INL RAVEN automates probabilistic risk assessment calculations that previously required months of expert analyst time. The nuclear engineer's work is shifting: routine neutronics screening and sensor surveillance are increasingly AI-assisted; the non-delegable human role concentrates on regulatory judgment, safety-case construction, and the licensed engineer-of-record accountability that no AI can assume.

  Effect on the work

  Early indications from SMR development programs suggest that GPU-accelerated simulation allows smaller engineering teams to complete design iterations that previously required larger ones, though the BLS 8% growth projection reflects the overall balance of SMR demand growth versus productivity gains. The net effect on headcount through 2034 remains contested.