Cavitation Assessment in Francis Turbines and Pumps-as-Turbines Using Machine Learning

Cavitation is an important factor limiting the performance and durability of hydraulic turbines. Computational fluid dynamics (CFD) provides access to local pressure and vapour-volume-fraction (VVF) fields, but its computational cost can limit the assessment of broad operating envelopes. This study develops and evaluates a CFD-conditioned Random Forest surrogate for voxel-level cavitation assessment in two hydraulically different machines: a pump-as-turbine and a Francis turbine. CFD-derived pressure and VVF fields were converted into a common voxel representation and used to define operational labels for cavitating and intensely cavitating regions. Several pressure- and VVF-based quantities were evaluated as candidate Stage-1 regression targets. The final Stage-1 representation retained three predicted VVF descriptors: voxel mean, 95th percentile and standard deviation. These predictions, together with spatial, operating-condition and neighbourhood features, were subsequently used by Stage 2 to classify cavitation occurrence and intensity. The final classifier does not require true voxel-level CFD descriptors at inference time. Complete CFD operating cases were held out during grouped validation to prevent leakage between spatially correlated voxels. The results show that the surrogate can reproduce the principal CFD-derived cavitation patterns for previously unseen operating cases within the represented machine-specific envelopes. However, strict zero-shot transfer between the two machines produced poor performance, demonstrating that the fitted model is not geometry-independent and requires target-machine adaptation. The proposed framework should therefore be interpreted as a rapid CFD-based screening tool for prioritising operating conditions and spatial regions requiring further numerical or experimental investigation, rather than as a replacement for detailed CFD or experimentally validated cavitation monitoring.