Evaluation of COBRA-TF prediction performance enhanced by machine learning-based CHF models

Abstract

This study investigates the integration of machine learning (ML) techniques into the thermal-hydraulic code COBRA-TF as a foundational advancement for predictive modeling in complex multiphase energy systems. Accurate prediction of critical heat flux (CHF) is essential not only for nuclear reactor safety and performance, but also for a wide range of high-efficiency energy systems that rely on advanced heat transfer, including concentrated solar power, supercritical CO₂ cycles, fission and fusion energy systems, and next-generation thermal desalination and storage technologies. Four ML models were developed and assessed, including two pure data-driven artificial neural networks (ANN) with 3- and 4-layer architectures, as well as two physics-informed machine learning (PIML) variants that embed physical constraints derived from the classical Zuber correlation. The models were trained and validated using the NRC CHF dataset and subsequently integrated into the COBRA-TF at the source-code level, enabling efficient and real-time hybrid modeling. Testing across diverse operating conditions demonstrated that all ML-enhanced COBRA-TF models substantially outperformed the legacy tool in both accuracy and stability. The PIML-enhanced models, particularly the PIML-4L model, achieve the lowest mean absolute error (MAE) and mean absolute percentage error (MAPE) among all tested models. While PIML-3L yields the lowest RMSE, PIML-4L performs best in normalized metrics such as rRMSE, indicating better control of extreme deviations. Error-based analysis further reveals that at least 80 % of predictions from all models fall within a ±15 % relative error range. PIML-3L achieves the highest accuracy under strict error tolerances (≤5 %), while PIML-4L performs best in the moderate error range (5–15 %), demonstrating superior robustness. Overall, it was concluded that integrating ML—especially PIML—into COBRA-TF significantly improves its CHF prediction capabilities, with PIML-4L offering the most comprehensive performance gain.