Saroj UpretiSelf-driving labs for materials discovery
Control room

Learning the processing–structure map

A machine learning framework that predicts block copolymer thin-film morphology from processing conditions, trained on high-throughput scattering data across six chemistries.

Headline result
6block copolymer chemistries in one generalizable model
Period
2024 – 2026
With
The University of Southern Mississippi · Advanced Light Source · NSLS-II
  • Machine learning
  • Structure–property relationships
  • Scattering analysis
  • Open database

The dataset came first

The robotic platform produced something the field mostly lacks: a large, consistent body of processing–structure data across multiple chemistries, failures included. Published data records what worked; uncurated high-throughput data also contains the disordered and dewetted films, which is what lets a model learn where the boundaries are.

What the model does

The framework predicts scattering descriptors — domain spacing, degree of order and morphological state — from composition, solvent, annealing protocol and film thickness. Predicting reduced descriptors rather than raw patterns keeps it interpretable and usable directly as a Bayesian optimization objective. Companion work on chemical-similarity-informed representations lets it transfer across chemistries rather than learn one system at a time.

Predicted against measured descriptors, with held-out test points in green. The scattering descriptors are predicted well — domain spacing R² 0.80, peak width 0.75 (random forest) — and the real-space AFM ones less so: domain spacing 0.54 (XGBoost), grain size 0.32 (SVM).Adapted from Lamb, Upreti et al., Macromolecules 59, 714 (2026), Fig. 7, under CC BY 4.0.

A model that has seen six chemistries gives a closed-loop optimizer an informed starting prior, and every autonomous campaign returns new training data. The data is published openly at polymermorphology.org.

Published in