Bridging micro-physical limits and macroeconomic shocks in SMRs: a physics-informed machine learning approach for LCOE prediction

This study presents a Physics-Informed Machine Learning (PIML) framework that integrates the IAEA ARIS database, international fuel test results, and deterministic nuclear physics codes with stochastic market variables to model the fuel cycle economics of Light Water Reactor (LWR)-based Small Modular Reactors (SMRs). Unlike static approaches, this methodology combines micro-level physicochemical and thermomechanical core degradations (such as fission gas release, hydrogen embrittlement, and fuel swelling) with macroeconomic market shocks within a unified analytical structure. The developed XGBoost model was trained on an 81,000-row matrix across five commercial SMR technologies (NuScale, VBER-300, BWRX-300, i-SMR, and ACP100) and validated using an independent dataset of 64,800 rows. The framework exhibited robust surrogate performance, achieving an R2 of 0.982 and an RMSE of 1.67 $/MWh. Multidimensional sensitivity analyses revealed that capacity factor reductions below 75% trigger a more substantial, exponential cost increase on the Levelized Cost of Electricity (LCOE) compared with uranium price fluctuations, escalating the LCOE well above 145 $/MWh. On the micro-physical scale, premature fuel discharge, necessitated when the hydrogen embrittlement threshold (>600 ppm) in the zirconium cladding is breached, completely neutralizes the 125–127 $/MWh optimal cost advantages achieved through extended burnup. This research provides a dynamic, risk-aware preliminary evaluation framework for investors and policymakers to optimize reactor selection and fleet planning.