Designing Reactors with AI: A New Core Approach

by

Author: Denis Avetisyan

Researchers have developed an artificial intelligence system capable of autonomously generating optimized nuclear reactor core designs, pushing beyond the limitations of conventional engineering.

ReactorFold reformulates nuclear fuel assembly design as a language modeling task by rasterizing lattice structures into token sequences, then employs a two-stage training pipeline-Base Full Fine-Tuning on low-fidelity data to establish geometric principles, followed by Low-Rank Adaptation on high-fidelity data to refine physical correlations-and finally aligns the model with multi-objective safety constraints-including $k_{eff}$, $F_q$, and $F_{\Delta H}$-through Direct Preference Optimization utilizing physics-based feedback from the OpenMC simulator.
ReactorFold reformulates nuclear fuel assembly design as a language modeling task by rasterizing lattice structures into token sequences, then employs a two-stage training pipeline-Base Full Fine-Tuning on low-fidelity data to establish geometric principles, followed by Low-Rank Adaptation on high-fidelity data to refine physical correlations-and finally aligns the model with multi-objective safety constraints-including $k_{eff}$, $F_q$, and $F_{\Delta H}$-through Direct Preference Optimization utilizing physics-based feedback from the OpenMC simulator.

ReactorFold leverages foundation models and Monte Carlo simulation to discover asymmetric reactor configurations and expand the possibilities for nuclear energy.

Designing optimized nuclear reactor cores presents a persistent challenge due to the vast and complex design spaces governed by intricate physical interactions. The work presented in ‘ReactorFold: Generative discovery of nuclear reactor cores via emergent physical reasoning’ introduces a novel generative framework leveraging large language models to autonomously explore and refine reactor core layouts. Remarkably, ReactorFold not only generates high-performing designs but also expands the conventional design space by independently adjusting key parameters and discovering asymmetric configurations inaccessible to traditional methods. Could this approach unlock fundamentally new reactor designs and reshape the future of nuclear energy innovation?

The Limits of Deterministic Design

Small Modular Reactor (SMR) core optimization has historically depended on deterministic frameworks – computational methods that meticulously solve the complex equations governing neutron transport and reactor physics. These approaches, while rigorously accurate, demand substantial computational resources and a deep understanding of nuclear engineering principles to effectively implement. Each fuel loading configuration and core design iteration requires extensive calculations to predict reactor performance, including power distribution, fuel burnup, and safety margins. The inherent complexity arises from the vast number of possible arrangements and the intricate interplay of physical phenomena within the reactor core, making even seemingly minor design changes computationally expensive and time-consuming. Consequently, traditional optimization methods often necessitate significant expert knowledge to guide the search for optimal designs and interpret the resulting data, creating a bottleneck in the innovation process.

Optimizing fuel loading and predicting reactor behavior present a significant computational challenge due to their inherent combinatorial complexity. Each potential fuel arrangement introduces a multitude of physical interactions that must be modeled with precision, yet the sheer number of possible arrangements quickly becomes unmanageable. Traditional methods, while accurate for individual scenarios, struggle to efficiently explore the vast design space – a space that grows exponentially with the size and complexity of the reactor core. This limitation hinders the identification of truly optimal designs and slows the pace of innovation, as each new configuration demands extensive computation and expert analysis. Consequently, the pursuit of improved reactor performance is often constrained by the practical difficulties of navigating this complex landscape, demanding alternative approaches capable of handling the combinatorial burden.

The ambitious goals of the Genesis Mission – specifically, achieving unprecedented levels of reactor performance and fuel utilization – necessitate a departure from traditional, computationally intensive design methodologies. Conventional optimization techniques, while rigorously tested, are proving inadequate to meet the accelerated timelines and demanding specifications of this new initiative. This push for innovation isn’t merely about faster processing; it demands fundamentally new approaches to reactor design, potentially leveraging machine learning, probabilistic modeling, and other advanced computational strategies to navigate the immense complexity of fuel loading and reactor physics. The Genesis Mission, therefore, serves as a catalyst, forcing a re-evaluation of established paradigms and opening doors to previously unexplored avenues in sustainable energy research.

The ReactorFold model demonstrates superior optimization efficiency and autonomously navigates beyond the reactivity and inventory constraints of a genetic algorithm baseline, expanding the design space and achieving improved peaking factor correlations.
The ReactorFold model demonstrates superior optimization efficiency and autonomously navigates beyond the reactivity and inventory constraints of a genetic algorithm baseline, expanding the design space and achieving improved peaking factor correlations.

Machine Learning: A Path Toward Efficient Core Design

Multiple machine learning techniques are being investigated to expedite Small Modular Reactor (SMR) design processes. Surrogate Models, often employing techniques like Gaussian Process Regression or Polynomial Chaos Expansion, create computationally inexpensive approximations of complex reactor physics simulations, enabling rapid evaluation of design parameters. Simultaneously, Convolutional Neural Networks (CNNs) demonstrate potential in analyzing spatial data inherent in reactor core layouts and neutron flux distributions, allowing for pattern recognition and predictive modeling of core performance. These methods aim to reduce the reliance on computationally intensive Monte Carlo simulations and other first-principles calculations, thereby decreasing design cycle times and associated costs. Initial applications focus on optimizing fuel assembly arrangements, predicting power distributions, and assessing core thermal-hydraulic behavior, though validation against high-fidelity simulations remains crucial.

A primary challenge in applying machine learning to reactor core optimization lies in the extensive data requirements for training accurate and reliable models. Traditional machine learning algorithms, such as surrogate models and convolutional neural networks, typically necessitate large datasets encompassing a wide range of operating conditions and core configurations to achieve acceptable performance. Furthermore, these models often exhibit limited generalization capabilities, meaning their predictive accuracy degrades when applied to scenarios outside the training dataset – for example, different fuel types, power levels, or core geometries. This lack of robustness can necessitate costly and time-consuming retraining or the development of specialized models for each unique reactor design or operating condition, hindering the broader adoption of machine learning in this field.

Deep Reinforcement Learning (DRL) and Physics-Informed Neural Networks (PINNs) represent advancements in applying machine learning to reactor core optimization by mitigating the data dependency and generalization challenges of traditional methods. DRL employs an agent that learns optimal control policies through trial and error within a simulated reactor environment, leveraging rewards to refine its actions and adapt to varying conditions without requiring extensive pre-labeled datasets. PINNs, conversely, integrate governing physics equations – such as neutron diffusion or heat transfer equations – directly into the neural network’s loss function. This constraint ensures that the network’s predictions adhere to known physical laws, improving accuracy and generalization capability, particularly in scenarios with limited experimental data. Both techniques aim to reduce reliance on large training datasets and enhance the robustness of machine learning models for complex reactor core designs.

ReactorFold: A Foundation Model for Core Design

ReactorFold represents a departure from traditional Small Modular Reactor (SMR) core design methodologies by leveraging the principles of Foundation Models and Large Language Models, an approach analogous to the successes observed in artificial intelligence programs such as AlphaGo and AlphaFold. These AI systems demonstrated the capacity to solve complex problems through pattern recognition and predictive analysis; ReactorFold aims to replicate this capability within the nuclear engineering domain. By adapting techniques proven effective in these unrelated fields, the project seeks to automate and optimize the design process, potentially leading to more efficient, safer, and cost-effective SMR configurations. The core innovation lies in framing core design as a problem solvable through the predictive power of large-scale AI models.

The representation of a Small Modular Reactor (SMR) fuel assembly lattice as a token sequence is central to the ReactorFold methodology. This discretization process transforms the physical arrangement of fuel rods, moderator, and structural components into a series of discrete tokens, analogous to words in a natural language. Each token encodes specific attributes of a lattice element, including material type, isotopic composition, and spatial coordinates. This tokenized representation allows the model to process core configurations as sequential data, leveraging the strengths of Large Language Models in pattern recognition and prediction. By treating lattice arrangements as ‘sentences’ composed of these tokens, ReactorFold can ‘understand’ core geometries and manipulate them to explore design alternatives and optimize performance characteristics.

ReactorFold leverages the Gemma 3 270M language model as its foundational architecture and employs data serialization to convert complex physical configurations of a Small Modular Reactor (SMR) core into a format suitable for processing by the model. This serialization process translates parameters defining the fuel assembly lattice – including isotopic compositions, geometric arrangements, and material properties – into a token sequence. Specifically, each physical characteristic is mapped to a unique numerical token, creating a discrete representation of the core’s state. This tokenized data then serves as the input for Gemma 3 270M, enabling it to learn relationships between core configurations and resulting reactor behavior without requiring direct manipulation of continuous physical parameters.

ReactorFold leverages both full fine-tuning and parameter-efficient adaptation techniques to achieve optimized core designs and behavioral predictions. Full fine-tuning involves updating all model parameters using a dataset of core configurations and corresponding reactor behavior, allowing the model to fully specialize to the domain. Complementing this, parameter-efficient adaptation methods, such as Low-Rank Adaptation (LoRA), minimize computational cost and data requirements by only training a small subset of parameters-specifically, low-rank matrices added to existing weights. This approach maintains performance comparable to full fine-tuning while significantly reducing training time and resource consumption, enabling rapid iteration and exploration of different core designs and operating conditions.

ReactorFold autonomously designed reactor core layouts, surpassing both genetic algorithm baselines and symmetric designs in total fitness, reactivity control, power peaking, and enthalpy rise, even while independently selecting a non-trivial, asymmetric fuel inventory.
ReactorFold autonomously designed reactor core layouts, surpassing both genetic algorithm baselines and symmetric designs in total fitness, reactivity control, power peaking, and enthalpy rise, even while independently selecting a non-trivial, asymmetric fuel inventory.

Validation and Future Implications: A New Era in Reactor Design

Rigorous validation underpins the development of ReactorFold, achieved through seamless integration with industry-standard reactor physics codes such as OpenMC and broader Monte Carlo simulation frameworks. This process ensures the model’s predictions align with established physics principles and accurately reflect real-world reactor behavior. By leveraging these trusted codes as a benchmark, researchers can confidently assess ReactorFold’s performance, verifying its ability to generate physically plausible and reliable designs. This validation strategy is critical for building trust in AI-driven nuclear engineering and facilitates the responsible deployment of innovative reactor designs.

The model’s ability to generate viable small modular reactor (SMR) designs stems from its implementation of Direct Preference Optimization (DPO). Unlike traditional optimization methods that rely on reward functions, DPO directly trains the foundation model to align with human-defined preferences and fundamental physics constraints. This approach bypasses the need for complex reward engineering, instead focusing on learning from comparisons – identifying designs that better satisfy critical parameters like neutronics and thermal-hydraulic stability. Consequently, the resulting designs are not merely optimal according to a mathematical function, but demonstrably realistic and physically plausible, offering a pathway to solutions that adhere to the stringent requirements of nuclear engineering and facilitating rapid exploration of the design space.

The methodology facilitates swift investigation of critical design variables, notably gadolinium (Gd) inventory-a key determinant of neutron absorption and fuel cycle length. By efficiently mapping the relationship between Gd concentration and reactor performance metrics, the approach unlocks opportunities to optimize fuel utilization and extend operational lifespan. This rapid exploration capability moves beyond traditional, computationally expensive methods, enabling engineers to identify superior fuel designs with fewer simulations. Consequently, the potential exists to significantly enhance reactor efficiency, reduce waste, and ultimately lower the cost of nuclear energy production, offering a pathway towards more sustainable and economically viable power generation.

The integration of foundation models into Small Modular Reactor (SMR) design represents a significant leap toward AI-driven innovation in nuclear energy, with designs such as the NuScale Power Module poised to benefit from these advancements. Recent work demonstrates that ReactorFold, a novel application of this technology, substantially outperforms traditional optimization methods; it achieved a six-fold improvement in design fitness compared to a genetic algorithm baseline. Crucially, this enhanced performance was realized within a computationally feasible framework, requiring only 1,000 high-fidelity simulations – a resource reduction that promises to accelerate the development cycle and unlock more efficient and effective reactor designs. This success suggests a future where AI not only assists in reactor design, but actively pioneers novel configurations previously unattainable through conventional methods.

The pursuit of optimized reactor core designs, as demonstrated by ReactorFold, exemplifies a dedication to elegant solutions. The system’s ability to bypass conventional, symmetric configurations and autonomously explore a broader design space speaks to a fundamental principle: complexity often obscures the most effective path. As John McCarthy once stated, “The best way to predict the future is to invent it.” ReactorFold doesn’t merely predict better designs; it actively invents them, illustrating that true innovation arises not from incremental improvements, but from a willingness to challenge established norms and embrace emergent possibilities. This aligns perfectly with the core idea of expanding beyond fixed constraints and discovering truly novel configurations.

Further Refinements

The demonstrated capacity for autonomous design, while notable, merely shifts the locus of complexity. The generative model excels at proposing configurations; rigorous validation remains stubbornly empirical. Future work must address the translation of emergent designs into physically realizable geometries, accounting for manufacturing tolerances and material constraints – the practicalities that presently limit exploration.

A pertinent question arises: does optimization, divorced from fundamental physical insight, yield truly superior designs, or simply novel arrangements within a constrained parameter space? The pursuit of asymmetry, though promising, demands careful consideration of stability and control mechanisms-areas where current simulations offer approximations, not certainties.

Ultimately, the value of this approach lies not in replacing human expertise, but in augmenting it. The model’s strength is breadth of exploration; the critical task remains discerning significance from the multitude of possibilities. A reduction in the cost of physical prototyping, coupled with increasingly accurate predictive models, may yet reveal if this expanded design space contains genuinely advantageous configurations, or merely a more elaborate echo of existing knowledge.

Original article: https://arxiv.org/pdf/2512.15756.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2025-12-21 06:46