Image-based estimation of component ratios in pharmaceutical powders using optical microscopy and deep learning

Ensuring the uniformity of excipient mixtures is a critical aspect of pharmaceutical manufacturing. In this study, we propose a non-destructive method for predicting the mixing ratios of excipients using optical microscopy images combined with convolutional neural networks (CNNs). Two CNN architectures, VGG16 and Xception, were trained to perform regression tasks on images of binary and ternary powder mixtures. Evaluation using test datasets demonstrated high prediction accuracy in both cases. For binary mixtures (CEOLUS PH-101 and PEARLITOL 200SD), the Xception model achieved a coefficient of determination (R 2 ) of 0.82, indicating a high goodness-of-fit, along with a low root mean squared error of prediction (RMSEP) and mean absolute error (MAE). For ternary mixtures (Ibuprofen, Pharmatose 100M, CEOLUS UF-702), the Xception model again outperformed VGG16, yielding lower RMSEP and MAE values and highlighting its superior predictive capability. Visualization of the learned feature space using UMAP revealed meaningful clustering patterns according to mixing ratios, indicating that the models effectively captured subtle morphological differences in particle structures. These results demonstrate that CNN-based models can accurately estimate mixing ratios from microscopy images alone. This approach offers significant advantages over conventional sampling-based analytical methods, particularly in scenarios requiring non-destructive monitoring or when sample quantities are limited. Furthermore, the ability to evaluate mixture uniformity directly from image data suggests strong potential for real-time process monitoring. Ultimately, the proposed technique may serve as a practical tool for quality control in pharmaceutical powder blending. • Deep learning predicts excipient ratios from simple microscope images. • Xception-based CNN outperforms VGG16 in estimating mixture compositions. • AI captures subtle particle features undetectable by human eyes. • UMAP reveals clear feature separation by mixing ratio. • A powerful, non-destructive tool for quality control in pharma manufacturing.