Abstract
<jats:p> Accurate computational modeling of the glass transition is a longstanding challenge in polymer chemistry. Atomistic molecular dynamics simulations offer a powerful approach to observe the changes in viscoelasticity, density and chain mobility associated with the glass transition, yet they often overestimate the glass transition temperature (T <jats:sub>g</jats:sub> ) and have limited sensitivity to number average molecular weight and dispersity. These deficiencies are largely attributed to high levels of noise in simulation data masking the T <jats:sub>g</jats:sub> , inaccurate polymer representations, and limited force field transferability. </jats:p> <jats:p> In this study, we build a robust and automated workflow to calculate T <jats:sub>g</jats:sub> using unsupervised machine learning, and apply it to a range of polymer models and dispersities. We show that our workflow provides improved T <jats:sub>g</jats:sub> estimates when compared to other established, 'bulk' measurements (e.g., monitoring changes in RMSD or density). We also benchmark different neural network charge models, which streamline polymer force field parameterization, and assess their influence on the T <jats:sub>g</jats:sub> . </jats:p>