ReactorFold: Generative discovery of nuclear reactor cores via emergent physical reasoning

ReactorFold: Generative discovery of nuclear reactor cores via emergent physical reasoning Yoonpyo Lee1* 1*Department of Nuclear Engineering, Hanyang University, 222, Wangsimni-ro, Seongdong-gu, Seoul, 04763, Republic of Korea. Corresponding author(s). E-mail(s): lukeyounpyo@hanyang.ac.kr; Abstract Designing nuclear reactor cores requires navigating large discrete design spaces governed by complex neutronic interactions. Traditional deterministic, meta- heuristic, and machine-learning-assisted methods search within fixed, human- defined configuration spaces, limiting their ability to discover fundamentally new design topologies. Here we introduce ReactorFold, a generative framework that reformulates fuel-assembly design as a sequence modeling problem for language models. Using Monte Carlo data, parameter-efficient fine-tuning, and Direct Preference Optimization (DPO), the model learns the latent structure of a pressurized-water-reactor assembly and generates candidate layouts in a single forward pass. Notably, the DPO-aligned model exhibits emergent design-space expansion: despite being trained exclusively on configurations with a fixed num- ber of gadolinium burnable absorber (Gd) rods, it autonomously adjusts Gd inventory to satisfy strict power-peaking constraints. The model also discovers high-performing asymmetric configurations that challenge conventional symmet- ric loading heuristics, accessing design regimes inaccessible to conventional search methods and demonstrating that language models can internalize causal physical relationships and transcend human-imposed design constraints. Keywords: Nuclear Reactor Design, Emergent AI Reasoning, Monte Carlo Simulation, Foundation Models, Physics-Aware Discovery 1 arXiv:2512.15756v1 [cs.LG] 12 Dec 2025 1 Introduction As the world strives to achieve carbon neutrality while meeting rapidly increasing electricity demands, nuclear energy has garnered renewed attention as a reliable, low- carbon baseload power source [1]. In particular, Small Modular Reactors (SMRs) are emerging as a pivotal solution to address energy security and grid flexibility challenges. Their modular design facilitates incremental deployment and integration into diverse energy markets [2], making the development of efficient and safe SMR core designs a critical priority. This operational urgency has been further amplified by the recent launch of the ”Genesis Mission,” a national directive that explicitly identifies nuclear fission as a priority challenge for artificial intelligence (AI)-accelerated innovation [3]. This policy pivot signals that the integration of AI into nuclear engineering is no longer merely an optional enhancement, but a strategic necessity to meet the accelerated deployment schedules required for global carbon neutrality. Historically, SMR core optimization has relied on deterministic and metaheuris- tic frameworks. Studies have applied algorithms like particle swarm optimization and genetic algorithms (GA) to refine various SMR concepts, including seed–blanket configurations, soluble-boron-free lattices, and fuel management strategies [4–10]. Additional research has explored fuel–material optimization for thorium-based and lead-based modular reactor cores [11, 12]. While effective, these methods typically depend on iterative high-fidelity simulations over user-defined search spaces, limiting scalability. To address these limitations, machine-learning (ML) assisted optimization has been extensively investigated. Deep reinforcement learning and convolutional neural net- work surrogates have been employed for combinatorial fuel loading in light-water and high-temperature reactors [13–15], while other studies utilized surrogates for back- end fuel cycle optimization, including canister loading and inverse depletion analysis [16–18]. ML models have also been integrated into advanced frameworks to enforce physics constraints, such as crud-aware optimization and multi-objective algorithms for coupled experiments [19, 20]. Furthermore, hybrid physics–ML approaches have improved core geometry optimization and parameter prediction [21, 22]. Physics- informed neural networks (PINNs) have emerged for embedding governing equations into reactor physics problems, including neutron diffusion and eigenvalue calculations [23–25]. Specifically for SMRs, surrogate-assisted GA have optimized dual-cooled fuel assemblies, reducing computational costs [26]. However, most of these schemes operate as forward or inverse solvers within an iterative loop, rather than directly generat- ing feasible layouts. This iterative paradigm confines exploration to human-predefined configuration spaces, as the search topology—such as fixed absorber inventory—must be specified a priori. The advent of large-scale foundation models (e.g., GPT-3, GPT-4, Gemini) has catalyzed interest in generative AI across scientific domains [27–29]. In nuclear engi- neering, large language models (LLMs) have primarily been e

This content is AI-processed based on open access ArXiv data.