Abstract
Predictive modeling of surfactant properties is constrained by limited experimental datasets, a common challenge in specialty chemical development where property measurements require specialized equipment and significant time investment. In this study, we address data scarcity through a semi-supervised generative augmentation framework that leverages both labeled and unlabeled molecular data for surface tension prediction. We implemented a two-stage variational autoencoder (VAE) training strategy using a curated database of 600 non-ionic surfactants. First, 461 unlabeled surfactant structures were used for VAE pre-training to learn latent representations capturing molecular connectivity patterns and amphiphilic relationships. Second, 125 molecules with surface tension measurements were used for fine-tuning to embed property-structure relationships. Our stratified generation framework produces surfactants matching target property distributions (Wasserstein distance = 0.030, KS statistic = 0.152) through stratified sampling and multi-criteria filtering, achieving 88.5% structural validity. We systematically evaluated augmentation effectiveness across seven levels (×1.25 to ×5) using graph convolutional networks with five random seeds per level. Results show optimal performance at ×3 augmentation (375 training molecules), where test R² improved from 0.60 (baseline, 125 molecules) to 0.78 (+28%), with RMSE decreasing to 0.056 mN/m. Multi-seed analysis demonstrates stable and reproducible training (test R² = 0.776 ± 0.030). This framework achieves performance equivalent to tripling the experimental dataset, providing a practical and scalable approach for specialty chemical property prediction where structural data is abundant but property labels are scarce.