Machine Learning of Methacrylated Dextran Phase Separation for Prediction, Transferability, and Adaptation
Year: 2026
Type: preprint
Abstract: Thermoresponsive polymers are promising materials for separations, sensing, drug delivery, and other biomedical applications, but their phase behavior depends on numerous polymer and solution variables, which creates a formulation space that is impractical to explore exhaustively by experiments. Here, we develop machine-learning (ML) models for predicting the lower critical solution temperature (LCST) of methacrylated dextran (Dex-MA), a biocompatible polysaccharide-based polymer, using 403 unique experimentally measured formulations spanning polymer molecular weight, degree of methacrylation, concentration, salts, surfactants, and other additives. The dataset is nonuniformly sampled because of practical synthesis constraints and application priorities. Despite this nonuniformity, descriptor-and molecular-embedding-based XGBoost and multilayer perceptron (MLP) models achieved mean absolute errors of 3.6 o C to 3.7 o C and 5.4 o C to 5.7 o C, respectively, for within-domain LCST prediction and mean phase-behavior classification accuracies of 89.4–89.6% and 85.3–86.6%, respectively. However, strong within-domain performance did not uniformly transfer to unseen formulation domains. Systematically withholding entire polymer, surfactant, and salt domains produced domain-dependent losses in predictive accuracy; for example, withholding 86 and 250 kDa Dex-MA increased LCST prediction errors to 11.4 o C and 17.5 o C, respectively, whereas errors for held-out surfactants were 5.6–7.2 o C. Progressive incorporation of target-domain measurements recovered predictive accuracy, but the recovery rate varied substantially with additive and salt identity and polymer molecular weight. This work provides a framework not only for predicting thermoresponsive-polymer phase behavior but also for evaluating model transferability to new formulation domains and guiding experimental data collection for model adaptation.
Source: ChemRxiv
Institutions: Louisiana State University, University of New Hampshire
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7221177562