Navigating Small Datasets with Machine Learning: Gaussian Process Modeling for Colloidal Gelation

Small datasets, which are common in many traditional science and engineering disciplines, present a challenge for powerful machine learning methods that are better suited for large datasets. Through the example of spontaneous colloidal sol-gel transition, we present a data-driven framework based on Gaussian Processes (GPs) for modeling and inference with limited data. GPs are well-suited for small datasets and automatically provide uncertainty quantification for model predictions. We explore GPs with two different structures (single and multi-output) to map synthesis parameters to gelation characteristics. While both GP structures quantitatively capture the relationships between the input and output variables, we favor multioutput GP, which attempts to simultaneously learn correlations between different outputs, due to its superior uncertainty calibration. To address inverse problems, we introduce a Bayesian framework that embeds a GP model into a Markov Chain Monte Carlo sampling algorithm. It enables the discovery of multiple synthesis pathways for achieving desired material properties. Therefore, this work presents a robust methodology for data-constrained modeling, with broad applicability to material design in general and the colloidal domain in particular.