Uncertainty Quantification in Machine Learning for Glass Transition Temperature Prediction of Polymers

---

Code repository for the above titled paper

Workflow

Dataset

• Dataset 1: 6097 homopolymers with Tg from PoLyInfo
• Dataset 2: 240 homopolymers with Tg from experiment data
• Dataset 3: 566 homopolymers with Tg from MD simulation
• High-Tg polymers (Tg>350℃): 19 high-Tg polymers from experiment data

Methods

• Neural network ensemble: Pytorch
• Gaussian process regression (GPR): GPy
• Monte Carlo dropout (MCD): Pytorch
• Mean-variance estimation (MVE): Pytorch
• Bayesian neural network (BNN): Pytorch
• Evidential deep learning (EDL): Pytorch, Chemprop

Input

• Morgan fingerprint with frequency: Considering the number of substructures

Output

• Mean and standard deviations of Tg for homopolyers

Metrics

• Spearman's rank correlation coefficient
• Calibration
• Sparsification