Flow regime identification of gas–liquid metal two–phase flow in a vertical pipe using differential pressure signals and unsupervised clustering

Accurate identification of flow regimes in gas–opaque liquid metal two–phase flow is crucial for the safe design and operation of LBE (Lead–Bismuth Eutectic)-cooled fast reactors. For high-temperature, opaque gas–liquid LBE two–phase flow, this study proposes and validates a non-intrusive flow regime identification method based on differential pressure fluctuation signals and unsupervised machine learning. Time-series differential pressure signals of gas–liquid two–phase flow were obtained experimentally in a vertical pipe with a diameter of 5 cm. Features were systematically extracted from the differential pressure signals using Power Spectral Density (PSD) analysis, forming a feature vector that includes the spectral centroid, peak frequency, sub-band energy ratios, spectral entropy, and other relevant features. The K-means clustering algorithm was employed to automatically partition the feature space, with the optimal number of clusters determined by the silhouette coefficient method. The study successfully identified four typical physical flow regimes: bubbly flow, cap bubbly flow, slug flow, and churn flow. The distribution of the clustering results on the void fraction-liquid superficial velocity plane aligns well with reference flow regime maps. Feature importance analysis revealed that PSD-derived features are the primary drivers of the clustering decision, with the top six features contributing to first Principle Component (PC1) loading all originating from PSD and accounting for a cumulative contribution exceeding 70%. This method uses low-frequency pressure fluctuations and K-means machine learning to overcome liquid metal opacity, enabling online flow regime identification in LBE loops.