ReactorFold

Language models discover non-intuitive reactor core designs through emergent physical reasoning

---

Overview

ReactorFold reformulates pressurized-water-reactor (PWR) fuel assembly design as a sequence modeling problem for language models. The model learns to generate candidate 17×17 fuel assembly layouts that satisfy neutronic constraints directly in the discrete design space.

Key Features

• Sequence-based reactor design: 2D fuel assembly → 1D token sequence
• Curriculum learning: Base FFT → LoRA fine-tuning → Online DPO
• Physics-in-the-loop: OpenMC Monte Carlo validation at every step
• Emergent behavior: Model autonomously expands Gd inventory beyond training distribution

---

Repository Structure

ReactorFold/
├── data_generation_100K.py    # Low-fidelity dataset (100k samples)
├── data_generation_10K.py     # High-fidelity dataset (10k samples)
├── sequence_length_check.py   # Token length validation
├── train_base.py              # Full fine-tuning on 100k dataset
├── train_lora.py              # LoRA fine-tuning on 10k dataset
├── dpo_10K_with_grid.py       # Online DPO with OpenMC-in-the-loop
├── ga_benchmark_hifi.py       # Genetic Algorithm baseline
└── symmetry_validation.py     # Symmetric benchmark generation

---

Installation

Requirements

• Python ≥ 3.10
• GPU with bfloat16 support (recommended)
• OpenMC with cross-section library

Python Dependencies

pip install numpy pandas tqdm datasets deap \
            torch transformers accelerate peft openmc

Environment Setup

import os
os.environ["OPENMC_CROSS_SECTIONS"] = "/path/to/your/cross_sections.xml"

---

Usage

1. Data Generation

Low-fidelity dataset (100k samples)

python data_generation_100K.py
# Output: reactor_100k_final.csv

High-fidelity dataset (10k samples)

python data_generation_10K.py
# Output: reactor_10k_final.csv

2. Model Training

Stage 1: Base model (Full Fine-Tuning)

python train_base.py
# Output: reactor_base_training_output/final_model/

Stage 2: LoRA Fine-Tuning

python train_lora.py
# Output: reactor_lora_training_output/final_model/

Stage 3: Online DPO

python dpo_10K_with_grid.py
# Output: dpo_results.csv, dpo_optimized_model/

3. Baselines

Genetic Algorithm

python ga_benchmark_hifi.py
# Output: ga_results.csv, ga_best_grid.txt

Symmetric Benchmarks

python symmetry_validation.py
# Output: symmetry_results_full.csv

---

Method

Training Pipeline

Stage	Data	Method	Output
1	100k low-fidelity	Full Fine-Tuning	Base model
2	10k high-fidelity	LoRA (r=32)	LoRA adapter
3	Online generation	DPO + OpenMC	Aligned model

Input Format

Reactor Core Design (k=1.05000, fq=1.0000, fdh=1.0000):
f f f g f f f f c f f f f g f f f
f f f f f f f f f f f f f f f f f
...

• f: Fuel rod (UO₂)
• g: Gadolinium burnable absorber (UO₂-Gd₂O₃)
• c: Guide tube / Instrument tube

Fitness Function

fitness = 0.6 × Fq + 0.4 × F_ΔH + P(k_eff)

where P(k_eff) penalizes deviations from target range [1.02, 1.08].

---

Results

ReactorFold achieves 6× improvement in fitness over GA baseline within the same computational budget (1,000 high-fidelity evaluations).

Method	Fitness	k_eff	Fq	F_ΔH	Gd Count
GA (fixed)	9.32	1.157	1.42	1.28	16
Sym-16	10.1	1.162	1.45	1.30	16
Sym-24	3.82	1.068	1.38	1.25	24
ReactorFold	1.55	1.052	1.32	1.21	29

---

Citation

@article{lee2025reactorfold,
  title={Language models discover non-intuitive reactor core designs 
         through emergent physical reasoning},
  author={Lee, Yoonpyo},
  journal={arXiv preprint},
  year={2025}
}

---

Data & Model Availability

• Dataset: Hugging Face
• Code: This repository

---

License

License information will be added later.

---

Contact

Yoonpyo Lee
Department of Nuclear Engineering, Hanyang University
📧 lukeyounpyo@hanyang.ac.kr