Intelligent Computational System for Constraint - First Governance of Subcritical Thorium - U233 Reactors

Title: Intelligent Computational System for Constraint - First Governance of Subcritical Thorium - U233 Reactors Type: Master of Technology (Department of CEAC, Specialization in Artificial Intelligence and Machine Learning) Final Dissertation | Researcher: Aurnab Dutta Roy (ORCID: 0009-0007-6258-0994) | Guide: Dr. Suryakanth V. Gangashetty (ORCID: 0000-0001-6745-4363) | Defense Date - 07 February, 2026 The Repository Contains Full Theory Paper, Reproducible Code, and generated datsets. Abstract: This work presents an Intelligent Computational System (ICS) designed to govern the design and operational safety of subcritical Thorium - 232 / Uranium - 233 reactor cores. Addressing the "trilemma" of fuel sustainability, waste management, and non - proliferation, this work introduces a "Constraint First" governance philosophy. Unlike traditional "performance first" optimization, the ICS prioritizes the enforcement of rigorous safety boundaries before evaluating performance metrics. The system utilizes a physics - grounded surrogate modelling approach, training an ensemble of XGBoost regressors on a dataset of 6,841 reactor configurations (generated via One - Group Diffusion Theory and augmented with synthetic data). Crucially, the ICS integrates Conformal Prediction to quantify epistemic uncertainty, converting point predictions into rigorous safety intervals calibrated for 90% marginal coverage. An Out-of-Distribution (OOD) detection module (Mahalanobis distance with D₉₅ threshold) acts as a veto mechanism, rejecting candidate designs beyond the model's validated domain to prevent overconfident extrapolation. Key Results: High - Fidelity Surrogate: The XGBoost-based surrogate achieves an R^2 > 0.99 for criticality k_eff and flux shape prediction, significantly outperforming Deep Learning baselines (FT - Transformer) on this tabular physics dataset. Rigorous Safety: The system demonstrates an empirical safety coverage of 93.3% on unseen test data, successfully identifying and rejecting unsafe supercritical designs Optimized Discovery: By filtering thousands of candidates through this safety gate, the ICS identified a stable reactor configuration capable of 8.7x higher breeding performance and 3.3x higher power output compared to a standard baseline, all while maintaining verified safety margins. This repository contains the full dissertation text, the complete Python codebase (Jupyter Notebooks), the generated datasets (Parquet), and the trained model artifacts, offering a reproducible framework for AI-driven nuclear safety research. Keywords: Nuclear Reactor Safety, Thorium Fuel Cycle, Uranium-233 Breeding, Artificial Intelligence, Machine Learning, Surrogate Modelling, XGBoost, Conformal Prediction, Uncertainty Quantification, Out – of - Distribution Detection, Constraint-First Governance, Subcritical Systems, IAEA, Human in the Loop AI Novelty Assessment Constraint - First Governance Logic: The concept of inverting the typical optimization structure, where safety constraints (criticality, power density) are not measured via penalties but "hard gates" applied via Conformal Prediction before any optimization occurs. Most nuclear AI literature treats safety as soft penalties in objective functions or post - hoc validation filters. Conformal Prediction in Nuclear Reactor Design: Application of split conformal prediction to generate distribution-free, finite-sample safety intervals (e.g., 93.3% empirical coverage for k_eff) for subcritical Thorium-U233 configurations. Standard uncertainty quantification (UQ) relies on Bayesian Neural Networks or Gaussian Processes, which require distributional assumptions and are computationally heavier for post-hoc calibration. Conformal prediction provides mathematical coverage guarantees without these constraints, making it particularly suited for real-time advisory systems under human governance. Out – of - Distribution Aware Assistant: Explicit implementation of an OOD detector (Mahalanobis distance with empirical D₉₅ threshold) as a veto mechanism for reactor design candidates outside the training distribution. This addresses a critical failure mode in AI - assisted safety - critical engineering: the system refuses predictions for geometries it has not reliably learned, rather than hallucinating plausible-but-unsafe recommendations. This aligns with IAEA guidance (SSG-52) that computational tools must explicitly bound their domain of validity. Strengths Scientific Rigour: The underlying physics (One - Group Diffusion with Bateman depletion) is implemented correctly and acknowledged with appropriate caveats (e.g., steady - state approximation). Methodological Integrity: The comparison between XGBoost and FT - Transformer is honest and scientifically valuable. Reporting the performances of both the XGBoost and the Transformer adds credibility and provides a useful benchmark for other researchers. Reproducibility: The complete code pipeline starting from physics based data generation (Chapter 5) through ensemble augmentation (Chapter 6) to ICS training and evaluation (Chapters 7 - 8) is provided as documented Jupyter notebooks. Trained model artifacts (XGBoost regressors, ensemble models) enable instant replication of all reported metrics without re-training. Human in the Loop Design: The system is framed as an advisory tool ("Intelligent Computational System"), not an autonomous controller, which aligns perfectly with nuclear regulatory standards (IAEA/ NRC). __________________ Code, Datasets and their relevance: 1)Chapter_5_Code Physics - Grounded Surrogate Dataset for Subcritical Thorium - U233 Reactor Governance Description: This repository contains the source code and generated datasets for Chapter 5 of the dissertation: "Intelligent Computational System for Constraint - First Governance of Subcritical Thorium - U233 Reactors." The code implements a "Physics - Grounded" data generation pipeline used to train surrogate models for reactor safety. It utilizes a One - Group Diffusion Theory approximation with Bateman depletion logic to simulate the steady-state physics of a Cylindrical Thorium - U233 core. Key Features: Methodology: Rejection sampling on random geometries (core radius R, height H, blanket thickness t, and initial U - 233 fraction). Safety Filter: The generator strictly rejects unsafe supercritical designs k_eff >= 0.98 and physically impossible states k_eff >= 1.0. It also rejects designs exceeding Power (500 MWth) or average core power density (100 W/cm³) envelopes. Physics Engine: Custom Python implementation of analytical flux solutions for cylindrical coordinates, coupled with ENDF/ B - VIII.0 nuclear data cross - sections. Output: A verified dataset of 1,500 subcritical reactor configurations, labelled with k_eff, Total Power, Power Density, and Breeding proxies (bred U-233 atoms after 1 day, plus a normalization by initial Th inventory). File Structure & Contents: 1. Source Code Step5_Dataset_Generation.ipynb: The primary Jupyter Notebook that runs the physics simulation. It handles: Anchoring nuclear constants to ENDF/B-VIII.0 values and saving them to nuclear_data.json for reproducibility (the run uses the in-memory nuc object). Defining the reactor geometry bounds (Radius: 50–150 cm, Height: 100–300 cm). Evaluating a closed - form one - group diffusion proxy with cylindrical buckling (fundamental mode) for 15,000+ random candidates. Filtering for the final 1,500 safe samples. 2. Configuration & Metadata dataset_schema.json: Defines the physical units, bounds, and descriptions for every column in the dataset (e.g., core_radius_cm, keff_est). material_library.json: Specifies the material compositions (Fuel: Th -U233 mix, Moderator: D2O, Reflector: Graphite) and number densities. nuclear_data.json: Contains the microscopic cross-sections σ_f, σ_c, ν derived from ENDF/B-VIII.0 for U - 233 and Th -232. exfor_cache.sqlite: Legacy Artifact. An empty SQLite database file. (Note: The initial design planned to fetch data dynamically from the IAEA EXFOR API, but the final implementation standardized on static ENDF values for reproducibility. This file is retained for code compatibility). 3. The Dataset surrogate_dataset.parquet: The primary dataset file (Apache Parquet format) containing the 1,500 samples. Recommended for use with Pandas (pd.read_parquet). surrogate_dataset.csv: A CSV version of the same dataset for human readability. surrogate_dataset.npz: A NumPy archive containing the data arrays for direct matrix manipulation. Usage Notes: To reproduce the dataset, open Step5_Dataset_Generation.ipynb and run all cells. The safety_margin column in the dataset represents the normalized distance to the nearest safety constraint (keff, power, power-density) Warning: This dataset is based on a steady-state diffusion approximation ("Gedanken" model) intended for AI governance research, not for licensing real - world reactor designs without high - fidelity validation (e.g., MCNP/OpenMC). 2)Chapter_6_Code Description: This repository contains the source code and augmented datasets for Chapter 6 of the dissertation: "Intelligent Computational System for Constraint - First Governance of Subcritical Thorium - U233 Reactors" To overcome the sparsity of the original physics simulated dataset (Chapter 5), this module implements an Epistemic Uncertainty Aware Data Expansion strategy. It utilizes a Deep Ensemble of 9 neural network regressors (scikit - learn MLPRegressor) to explore the reactor design space and generate synthetic samples. Crucially, it filters these samples based on prediction variance, ensuring that only high - confidence, physically valid designs are added to the training pool. Key Methodology: 1. Candidate Generation: Uses Latin Hypercube Sampling (LHS) / Random Uniform Sampling to propose 25000 theoretical reactor geometries within the design bounds. 2. De