Foundation language model for generative polymer design
POLYT5
An encoder-decoder chemical language model that predicts polymer properties and generates new polymer structures conditioned on target performance.
Read moreMaterials informatics / polymer ML / computational chemistry
Harikrishna Sahu
I build computational and machine-learning workflows that accelerate materials discovery, with recent work spanning polymer foundation models, physics-informed Gaussian Process Regression, autonomous structure generation, and data-driven design.
Focus
Research and engineering themes
Physics-informed Gaussian Process Regression for materials informatics
matgpr
A public Python package for reproducible materials-informatics workflows with Gaussian Process Regression, uncertainty-aware prediction, materials fingerprints, and physics-informed mean functions.
Read moreAutonomous atomic-scale polymer model generation
Polymer Structure Predictor
A Python toolkit that builds a hierarchy of polymer models from repeat-unit SMILES, including oligomers, infinite chains, crystals, and amorphous structures.
Read moreSelected work
Publications
On-the-Fly Machine-Learned Force Fields for High-Fidelity Polymer Glass Transition Simulations
Journal of Chemical Information and Modeling, 2026, 66(17), 11129-11139
Load-dependent hardness prediction for materials using machine learning
Applied Physics A, 2026, 132, 744
POLYT5: an encoder-decoder foundation chemical language model for generative polymer design
npj Artificial Intelligence, 2026, 2, 30
From Corpus to Innovation: Advancing Organic Solar Cell Design with Large Language Models
npj Computational Materials, 2026, 12, 27
polyBART: A Chemical Linguist for Polymer Property Prediction and Generative Design
Findings of ACL: EMNLP 2025, 2025, 12104-12119
Latest
Writing
What T5 Adds Beyond Decoder-Only Polymer Models
A technical but accessible note on why encoder-decoder T5 architectures matter for polymer property prediction, conditional generation, molecule-text translation, and materials discovery workflows.
Hardness Prediction Needs the Load
A note on our Applied Physics A paper showing why load-dependent experimental data can matter more than DFT-derived hardness proxies for Vickers hardness prediction.
Why I Built matgpr for Materials Informatics
A first note on matgpr, an open-source toolkit for uncertainty-aware, physics-informed Gaussian Process Regression workflows in materials informatics.