Accurate characterization of lithofacies and fault networks is critical for delineating reservoir architecture, heterogeneity, and compartmentalization in structurally complex hydrocarbon systems. Conventional seismic interpretation methods are often limited by subjectivity and insufficient resolution in complex stratigraphic and structural settings. This study evaluates the effectiveness of an integrated machine learning workflow combining Self-Organizing Maps (SOMs) and Convolutional Neural Networks (CNNs) for seismic lithofacies classification and fault detection in the Penobscot Field on the Scotian Shelf, Eastern Canada. A suite of multi-attribute seismic data derived from 3D post-stack seismic volumes was optimized using principal component analysis and multi-metric dependency assessment prior to SOM training. An 8 × 8 SOM grid was employed to classify seismic facies into 8 lithological classes, calibrated against petrophysical logs and core descriptions from 2 wells. The SOM achieved a high recall (1.00) across key reservoir facies, with precision values ranging from 0.50 to 1.00, reflecting effective discrimination of reservoir and non-reservoir lithologies despite transitional facies mixing. SOM training exhibited stable convergence, with neuron weight shifts stabilizing below 0.1 after approximately 40 training epochs, indicating robust learning of the seismic attribute space. Fault detection was performed using a pre-trained CNN to generate fault probability volumes, which were subsequently refined using Laplacian-of-Gaussian filtering and skeletonization. Integration of CNN-derived structural attributes within the SOM framework enabled near-sample-scale fault delineation (≈8 m vertical resolution) and the isolation of structurally coherent fault planes. The combined SOM-CNN approach enhanced the interpretability of stratigraphic architecture, fault compartmentalization, and hydrocarbon anomaly distribution within Jurassic-Cretaceous reservoir intervals, particularly the Missisauga and Abenaki formations. The results demonstrate that unsupervised learning, when supported by quantitative validation and structural constraints, provides a robust and interpretable framework for high-resolution reservoir characterization in complex offshore settings.