A Unified Experimental Database and Group-Aware Pilot Machine Learning for Flow Boiling in Rectangular Minichannels

A unified experimental database and a documented Python (3.12.6)-based harmonisation workflow are presented for local heat-transfer analysis in rectangular-minichannel flow boiling. The database integrates 449 experimental source files and 64,385,791 point-level records covering six working fluids, multiple surface conditions, channel configurations, and orientations. The workflow combines template-based files, central-line infrared wall-temperature data, geometric and operating metadata, local pressure and saturation temperature reconstruction, bulk fluid temperature interpolation, heat-loss correction, local heat-transfer coefficient and Nusselt number calculation, operational branch labelling, and auditable quality-control flags. The database is characterised at the point and source-file levels to identify branch imbalance, unequal source-file sizes, and heterogeneous coverage. A source-file-grouped pilot benchmark is conducted on streaming-sampled subcooled and saturated subsets using Random Forest regressors and five group-based train/test partitions. Across the five splits, the nine-predictor RF achieved R2 = 0.887 ± 0.015 for subcooled Nu and R2 = 0.810 ± 0.043 for the subcooled heat-transfer coefficient; saturated performance was weaker and more split-sensitive (R2 = 0.340 ± 0.134 for Nu and 0.317 ± 0.091 for the heat-transfer coefficient). These results are treated as feasibility screening rather than final model ranking or evidence of campaign- or configuration-independent transfer. The database architecture and group-aware pilot validation provide a documented foundation for controlled physical interpretation and further validation.