\title{\textbf{Machine Learning Surrogate Modelling for Rapid Steady-State Thermal-Hydraulic Safety Analysis of a 10 MW Material Testing Reactor}}

\title{\textbf{Machine Learning Surrogate Modelling for Rapid Steady-State Thermal-Hydraulic Safety Analysis\\ of a 10 MW Material Testing Reactor}} \author{} \date{} \begin{document} \maketitle \begin{abstract} Reliable steady-state thermal-hydraulic analysis is essential for evaluating the safety margins of a 10 MW Material Testing Reactor (MTR). High-fidelity simulation codes provide detailed predictions of reactor thermal-hydraulic behaviour, but repeated simulations can become computationally expensive when large numbers of operating conditions are considered for sensitivity analysis, uncertainty studies, and rapid safety assessment. This study develops a machine learning surrogate model and interactive user interface for rapid prediction of key thermal-hydraulic temperatures and safety ratios of a 10 MW MTR. A database of 1,000 steady-state simulations was generated using PLTEMP/ANL v4.3. Five input parameters were varied, namely channel thickness, system pressure, reactor power, inlet coolant temperature, and coolant mass flow rate. The resulting database contained four principal output parameters: minimum Critical Heat Flux Ratio (CHFR${\min}$), minimum Onset of Nucleate Boiling Ratio (ONBR${\min}$), maximum fuel temperature (T${\max,\mathrm{fuel}}$), and maximum cladding temperature (T${\max,\mathrm{clad}}$). Latin Hypercube Sampling was used to generate the simulation cases and provide broad coverage of the input space. The dataset was divided into 800 training samples and 200 testing samples. Random Forest, Gradient Boosting, and Feedforward Neural Network models were subsequently trained and evaluated to identify the most suitable surrogate modelling approach. The results show that all three machine learning approaches were able to learn the relationship between the reactor operating conditions and the thermal-hydraulic responses to varying degrees. Gradient Boosting achieved the highest predictive accuracy across the four output parameters, while Random Forest also demonstrated strong predictive performance. The Feedforward Neural Network produced comparatively poorer predictions. The actual-versus-predicted results for the Gradient Boosting model showed close agreement between the surrogate predictions and the PLTEMP/ANL simulation results, with the predicted values closely following the perfect prediction line for CHFR${\min}$, ONBR${\min}$, T${\max,\mathrm{fuel}}$, and T${\max,\mathrm{clad}}$. Sensitivity analysis identified coolant mass flow rate as the dominant input parameter controlling the predicted thermal-hydraulic responses. Its influence was particularly strong across the four output parameters, indicating that coolant flow conditions play a major role in determining reactor temperature behaviour and safety margins. The selected Gradient Boosting model was integrated into an interactive user interface that allows users to enter reactor operating conditions and obtain rapid predictions of thermal-hydraulic temperatures and safety indicators without executing a new PLTEMP/ANL simulation. The results demonstrate that machine learning surrogate modelling can effectively approximate high-fidelity steady-state thermal-hydraulic simulations while substantially reducing the computational effort required for repeated predictions. The combination of the Gradient Boosting surrogate model and interactive interface provides a practical framework for rapid reactor safety assessment and improves accessibility to thermal-hydraulic analysis. The developed approach can be extended to larger simulation databases, uncertainty quantification, transient analysis, and real-time reactor monitoring applications. \end{abstract} \noindent\textbf{Keywords:} Material Testing Reactor, Thermal-Hydraulics, PLTEMP/ANL, Machine Learning, Gradient Boosting, Surrogate Model, CHFR, ONBR, Reactor Safety. \end{document}