<p>Accurate mineral classification is critical for various geology, mining, and materials science applications. This study investigates the use of advanced deep learning architectures, such as EfficientNet and ResNet models, for automated mineral classification. Six architectures were evaluated to enhance performance, employing data augmentation and transfer learning (EfficientNetB0, EfficientNetB7, ResNet50, and ResNet152). The models were assessed based on classification precision, recall, validation accuracy, and F1 scores. Among the tested architectures, EfficientNetB7 and ResNet152 demonstrated superior performance, achieving the highest classification accuracies and robust F1 scores, particularly for complex mineral types such as quartz and malachite. EfficientNetB7 recorded the robust F1 score of 0.1450 compared to other architectures, while ResNet152 provided the most balanced performance across all mineral classes. Strategic learning rate reductions and architectural depth were crucial in mitigating overfitting, resulting in high stability and strong generalization capabilities. Despite their strong performance, specific architectures struggled to classify underrepresented mineral classes like muscovite, indicating class imbalance and feature complexity challenges. The results suggest that deeper architectures, combined with effective transfer learning and optimization strategies, are highly suitable for fine-grained mineral classification tasks. This research highlights the potential of deep learning in mineralogy and identifies avenues for future exploration, such as class balancing and combining ensemble methods to enhance model performance further.</p>