<p>Seismic reservoir characterization and geomechanical modelling rely on shear wave velocity (Vs), which is often costly to directly measure using dipole sonic logging, providing gaps in legacy well databases. A data-driven alternative may adopt a machine learning approach, but the current validation paradigm, such as random train-test splitting within a single well, exploits spatial autocorrelation rather than learning transferable petrophysical relationships and systematically overstates reported performance, creating models that do not generalize across wells. This study fills this generalization gap by developing a geology-conditioned composite ensemble model, which was tested under robust cross-well blind-test conditions in the complex Neogene clastic succession of the Suma Basin, Bangladesh. The workflow combines three complementary components: a distributional domain-adaptation step that ensures that inter-well feature drift is removed; a gamma-ray-based lithofacies classifier that divides the prediction space into sandstone, heterolithic, and shale domains; and a composite ensemble that routes a prediction to the algorithm most suitable to each inductive problem, for instance, Random Forest on high-variance gas-bearing zones. No data from blind wells were ever used in any training or optimization. The composite model was more robust than all the individual algorithms under operationally relevant interval-level accuracy under blind-well evaluation. Interestingly, the classical Castagna empirical correlation had the best goodness-of-fit across the world, highlighting the enduring usefulness of physics-based baselines to compaction-coherent siliciclastic successions. Such findings are viewed as evidence of concept: to be transferable to other Surma Basin wells or similar deltaic environments, larger multi-well datasets need to be validated.</p>