In recent years, the advent of Convolutional Neural Networks (CNNs) has opened up new avenues for advancing Surface Roughness (SR) prediction methodologies, particularly through the analysis of Scanning Electron Microscope (SEM) images. However, notable gaps existed in the literature regarding the application of CNNs to SEM images for SR prediction. This research addresses these existing gaps by employing a CNN to analyze SEM images for SR prediction, with a particular focus on comparative analysis of different magnification levels. Three distinct datasets, magnified at 150X, 250X, and 500X, were utilized, comprising 2097, 2103, and 2102 images, respectively. These images undergo preprocessing techniques to enhance the CNN model's ability to generalize to new images. Subsequently, a sequential CNN model, comprising 27 layers including convolutional, max pooling, batch normalization, flatten, and fully connected dense layers, is developed and trained on the datasets. The study provides detailed comparative analyses of accuracy, precision, recall, and F1-score across magnification levels. Results indicate that the dataset magnified at 500X consistently outperforms the others, exhibiting superior accuracy (75.7%), precision (0.65), recall (0.72), and F1-score (0.72). Higher magnification levels provide finer details and more explicit images, enabling the model to discern subtle features with increased accuracy. Additionally, the 500X dataset exhibits a better balance between minimizing false positives and false negatives, making it more suitable for real-world applications requiring detailed analysis of microscopic structures. These findings underscore the importance of selecting appropriate magnification levels in SEM imaging for accurate SR prediction.