Highlights• Generalized supervised machine learning models resulted in trustable synthetic well-logs for NMR porosity with a difference of less than 5% from real data.• We find improvements in total porosity estimation using machine learning models, in comparison to conventional petrophysical estimation.• For the first time, an explained machine learning model reveals the contributions of multiple features in a tree boosted algorithm using NMR well-log data.• We reveal the linear and nonlinear relationships in the contribution of features to machine learning models.