Abstract We utilize 3D seismic data and robust rock-physics models, combined with a well dataset, to investigate the subsurface of the Mundaú sub-basin, Brazil. Seismic attributes analysis and unsupervised machine-learning approaches were able to produce high-resolution images to allow the mapping of the 3D geometry of ancient geomorphological features across stratigraphic levels, from the Albian to the Turonian interval. Significant deep-water elements were identified using seismic attributes and machine-learning techniques (i.e. channel complex, point bars, feeder channels, faults, depocentres, dendritic lobes, smaller channels and distributaries). In addition, the petrophysical analysis enhanced the subsurface characterization by employing a deep convolutional network that allowed S-wave modelling and synthetic seismic generation. The well-log data analysis validated interpretation of sand-prone deposits; in addition, the rock-physics modelling provided insight into the deposited lithologies. After the petrophysical analysis, seismic facies classification was performed using machine-learning techniques, including self-organizing maps and independent component analysis, which provided valuable insights into the geomorphology of this under-researched basin. The enhancement of seismic and petrophysical data with machine learning proves to be a useful technique for better characterizing this basin. This approach may be used on similar frontier hydrocarbon basins to help de-risk petroleum exploration.