Automated characterization of nanomaterial structures from existing datasets is crucial for maximizing prior research utility and accelerating nanotechnology advancements. While scanning electron microscopy (SEM) is vital for nanomaterial visualization, manual SEM image analysis is time-consuming and subjective, hindering efficient data re-analysis and knowledge extraction. This study addresses automated nanomaterial characterization by developing deep learning models to classify geometric patterns of palladium nanoparticles (Pd NPs) in SEM images, using a pre-existing labeled dataset publicly available on Figshare. The dataset comprises SEM images of Pd NPs on carbon surfaces, categorized into ordered and disordered patterns from a previous experimental study. The ordered subset is further classified into five geometric pattern classes: bright large particles, circles, edges, lamellar curvatures, and single particles. We evaluated two deep learning approaches: a custom Convolutional Neural Network (CNN) and a hybrid of CNN and VGG16 model combining custom layers with a pre-trained feature extractor. The hybrid model achieved exceptional performance, with 100% accuracy score across all classes on the test dataset. The custom CNN model, in contrast, achieved 78% overall accuracy with recall limitations in certain classes. The hybrid model's superior performance highlights the effectiveness of transfer learning for enhancing classification accuracy with reduced model complexity. These findings demonstrate that the hybrid of CNN and VGG16 model provides a robust and efficient solution for automated nanoparticle geometric pattern classification in pre-existing SEM datasets. This approach enables accelerated and objective re-analysis of archived nanomaterial characterization data, streamlining workflows and maximizing insights from previously collected data.