Synthetic images from generative AI for compositional analysis of dried solution patterns

Image-based identification of solutes from dried drop stains offers a low-cost, scalable alternative to traditional chemical analysis, enabled by artificial intelligence and annotated image libraries. While recent advances in robotic droplet imaging have facilitated the collection of such datasets, experimental acquisition remains a key bottleneck. Here, we explore whether synthetic image generation can supplement or replace real data in the context of salt classification. Using Stable Diffusion with low-rank adaptation, we generate 133 000 synthetic images from as few as N = 1–128 experimental images per salt across seven common inorganic solutes. Synthetic images become visually indistinguishable from real ones for N ≥ 6; however, quantitative analysis based on 47 image metrics reveals subtle differences that vanish around N = 128. When used to train random forest, XGBoost, and multilayer perceptron (MLP) models, synthetic data alone achieve classification accuracies of up to 90% in some low-data regimes. Synthetic augmentation offers a powerful strategy for expanding training sets and enabling image-based chemical classification in data-scarce scenarios, even if it seldom outperforms models trained on real data. The realism of synthetic images also highlights the growing need for tools that detect AI-generated scientific images to ensure data integrity.