Reliability-assessed machine learning for high flux flow regime classification

Thermal-hydraulic systems in nuclear engineering require reliable monitoring, accurate characterization and models capable of handling nonlinear, time-dependent behavior. High-void-fraction multiphase flows introduce significant diagnostic challenges, as the chaotic, nonlinear nature of interfacial transitions makes conventional modeling and rigid thresholding approaches less practical. This study develops a diagnostic framework for classifying annular gas–liquid flow sub-regimes using centerline conductivity probe measurements. Continuous experimental runs were partitioned into short temporal segments, with feature vectors constructed from volumetric, frequency-domain and morphological characteristics to represent the geometric structure of the flow. Four supervised learning architectures, K-nearest neighbors (KNN), support vector machines (SVM), random forest (RF) and gradient boosted trees (GBT), were evaluated. Feature selection reduced the initial six-feature configuration to an optimal three-feature subset, improving classification performance across most architectures. The RF model achieved the best overall results, with a macro F1 score of 0.9263 at the local segment level and perfect balanced accuracy under global run-level majority voting. Probabilistic evaluation using expected calibration error (ECE), maximum calibration error (MCE) and Brier scores confirmed that the RF model produced the most reliable confidence estimates among all evaluated classifiers. Overall, the study shows that machine learning can be applied to flow characterization problems to address challenges that traditional methods handle poorly. The results also highlight the importance of understanding model behavior, not just predictive accuracy, through analysis of evaluation methodology and calibration of predicted probabilities.