Similarity-Based\nMachine Learning for Small Data Sets:\nPredicting Biolubricant Base Oil Viscosities

Machine learning (ML) has been successfully applied to\nlearn patterns\nin experimental chemical data to predict molecular properties. However,\nexperimental data can be time-consuming and expensive to obtain and,\nas a result, it is often scarce. Several ML methods face challenges\nwhen trained with limited data. Here, we introduce a similarity-based\nmachine learning approach that enables precise model training on small\ndata sets while requiring fewer features and enhancing prediction\naccuracy. We group molecules with similar structures, represented\nby molecular fingerprints, and use these groups to train separate\nML models for each group. We first validate our method on larger data\nsets of dynamic viscosity and aqueous solubility, demonstrating comparable\nor better performance than traditional approaches while requiring\nfewer features. We then apply the validated methodology to predict\nthe kinematic viscosity of biolubricant base oil molecules at 40 °C\n(KV40), where experimental data is particularly limited. Our method\nshows noticeable model performance improvement for KV40 prediction\ncompared to transfer learning and the standard Random Forest. This\napproach provides a robust framework for limited data that can be\nreadily generalized to a diverse range of molecular data sets especially\nwhen clear structural patterns exist in the data set.