This preprint applies neural-network and random-forest regressors trained on Rapster-generated synthetic cluster catalogs to map observable properties like total mass and half-mass radius directly onto the mass of the heaviest black hole formed via repeated mergers. The methodology relies on large ensembles of simulated clusters rather than direct observations, with no explicit sample size reported beyond the underlying Monte-Carlo runs; results therefore inherit all modeling assumptions in Rapster, including stellar-wind prescriptions and dynamical encounter rates. When the trained models are applied to real globular clusters, predicted IMBH occupation fractions fall to ~0.02 and maximum masses remain below ~100 solar masses, consistent with the absence of convincing IMBH detections in Milky Way globulars. Nuclear star clusters such as NGC 5102 and NGC 5206 emerge as more promising hosts, yet even there the forecasts lie below some kinematically claimed values, implying that gas or stellar accretion must supplement hierarchical mergers. The work notably omits detailed comparison with LIGO/Virgo events in the upper mass gap (e.g., GW190521) and does not quantify how binary-single interactions or primordial binaries alter runaway-collision probabilities. A normalizing-flow module is used to assess the likelihood of early collisional runaways, but this step remains sensitive to the initial mass function and cluster density profiles chosen in the training set. As an unreviewed arXiv submission dated May 2026, the findings should be treated as exploratory until independent N-body validation appears.