The rise of generative models has highlighted the importance of cross-domain applications with mixed data. Recent studies on learning intermodal representations have predominantly relied on supervised deep learning models, while unsupervised models play a secondary role in auxiliary tasks. This article proposes a new fully unsupervised approach to learning intermodal representations based on a topologically coherent map that allows bidirectional prediction/regeneration between domains. The method is evaluated on an unsolved problem in petrophysics: generating a complete set of basic logs from special acoustic image logs of wells in highly heterogeneous carbonate reservoirs in the Brazilian pre-salt. In addition, a supervised deep learning model was developed as a benchmark to evaluate the performance of our approach.