The Oil and Gas Reservoirs segment acquires large volumes of data to support decision-making, modeling, and optimizing the extraction of hydrocarbons. However, traditional analysis methods, which rely on classical mathematical and statistical models, often restrict the ability to extract valuable information from these data. Given the high acquisition costs and the complexity of the data, advanced machine learning techniques, such as Representation Learning, Generative Learning, and Regeneration Learning, emerge as promising alternatives to maximize the utilization of the information contained in the data. This thesis proposes a novel self-supervised methodology for bidirectional intermodal regenerative learning using Kohonen Self-Organizing Maps (SOM). The goal is to enable the association, generation, and regeneration of data between distinct modalities, exploring non-linear interrelationships and handling potentially incomplete data. The methodology is based on a conceptual adaptation of the SOM, transforming it into a heteroassociative memory for mappings between source and target data spaces with different modalities. This approach allows: (1) associating representations of distinct modalities in the same SOM, (2) generating data of lower complexity from more complex data, (3) regenerating complex data from simple data, and (4) handling missing values. Validation was performed in two previously unexplored applications of bidirectional data generation/regeneration in Reservoir Petrophysics. Type-I application involves the multivariate prediction of conventional logs from special acoustic image logs, and Type-II application involves the regeneration of acoustic image logs from a set of conventional logs. The analysis drew upon real data from wells drilled in highly heterogeneous carbonate reservoirs in the Brazilian pre-salt. The methodology demonstrated efficacy in bidirectional data generation and regeneration, enabling valuable information to be extracted from complex and potentially incomplete data. The results obtained highlighted the ability of the SOM to explore interrelationships between modalities and regenerate coherent data, even in challenging scenarios. The methodology proved capable of handling data with missing values, including the ability to regenerate images (Type-II application) from a very limited number of conventional logs. A further outcome of the thesis is a set of evaluation metrics for the images obtained in Type-II application, which are intuitive for a petrophysical interpreter. This thesis presents a significant contribution to the field by enabling the leveraging of information contained in the various types of data acquired in the Oil and Gas industry. The self-supervised SOM-based methodology proved effective for Regenerative Learning applications, offering potential to reduce costs, accelerate real-time decision-making, and increase the efficiency of data utilization and integration. Furthermore, the methodology can be extended to other intermodal mapping contexts, broadening its impact across different areas of knowledge.