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

TL;DR

ReactorFold reframes nuclear reactor core design as a sequence-generation problem and demonstrates that a language-model–driven generator can discover feasible, high-performing, and asymmetric lattice topologies beyond traditional symmetry and fixed-parameter constraints. The method uses a three-stage curriculum (FFT, LoRA, DPO) with Monte Carlo-based physics feedback to produce designs that are validated by OpenMC, achieving a six-fold improvement in objective fitness over a fixed-inventory GA within 1,000 high-fidelity evaluations. Key insights include emergent constraint relaxation (modulating Gd inventory) and the discovery of design-space regions inaccessible to conventional search, supported by analyses showing decoupled objective correlations. The work suggests a new paradigm for accelerated, physics-informed inverse design in nuclear engineering and potentially other discrete-design domains, while noting limitations to 2D cores and the need for multi-physics integration in future work.

Abstract

Designing nuclear reactor cores requires navigating large discrete design spaces governed by complex neutronic interactions. Traditional deterministic, metaheuristic, 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 number 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 symmetric 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.

Figures (3)

• Figure 1: Overview of the ReactorFold framework.a Data serialization strategy. The two-dimensional fuel assembly lattice is rasterized into a one-dimensional discrete token sequence to reformulate the design problem as a language modeling task. b Curriculum-based training pipeline. The model undergoes Base Full Fine-Tuning (FFT) on a large corpus of low-fidelity layouts to acquire geometric syntax, followed by Low-Rank Adaptation (LoRA) on high-fidelity data to refine physical correlations. c Alignment via Direct Preference Optimization (DPO). The model generates candidate layouts which are evaluated by the OpenMC simulator. The DPO algorithm uses these physics-based preferences to align the model with multi-objective safety constraints ($k_{\text{eff}}$, $F_q$, and $F_{\Delta H}$), effectively closing the physics feedback loop.
• Figure 2: Emergent design dynamics and strategy.a Optimization efficiency comparison showing the rapid convergence of the ReactorFold model compared to the GA baseline. b Shift in the Pareto frontier. The ReactorFold model autonomously breaches the reactivity barrier ($k_{\text{eff}} \approx 1.157$) that constrains the fixed-inventory GA baseline, navigating towards the ideal target zone. c Scatter plot of Gd inventory vs reactivity, highlighting the model's capability to expand the design space beyond the fixed inventory constraint. d Evolution of peaking factors ($F_q$ and $F_{\Delta H}$). e, f Correlation matrices for DPO and GA, respectively.
• Figure 3: Comprehensive performance analysis of the best-found designs.a--c Core maps of the best symmetric benchmarks with fixed Gd inventories (16, 24, 32). d The best layout found by the GA baseline (Gd=16). e The best layout generated by ReactorFold model. The model autonomously selected an inventory (Gd=29) and discovered a non-trivial, asymmetric pattern. f--i Quantitative comparison of total fitness, reactivity control ($k_{\text{eff}}$), power peaking factor ($F_q$), and enthalpy rise factor ($F_{\Delta H}$). The ReactorFold model consistently outperforms both GA and symmetric benchmarks.