Machine Learning-Accelerated Inverse Design of Sequence-Controlled Terpolymers
- 1Chemical Sciences Division
- 2Oak Ridge National Laboratory
Abstract
Sequence-controlled polymers hold immense potential for advanced materials design, yet discovering their experimental recipes is bottlenecked by the high-dimension kinetic parameter space of multicomponent copolymerization. Traditional ordinary differential equation (ODE) solvers provide rapid macrokinetic predictions but are fundamentally blind to explicit backbone microstructures. To bridge this gap, we use a stochastic simulation algorithm (SSA) engine to generate high-throughput terpolymer sequence data from a physically constrained parametric sweep across a 9-dimensional kinetic space, charting how feed stoichiometry and kinetic propensities dictate macromolecular structure. Information-theoretic analysis and t-SNE dimensionality reduction of the resulting 27-component triad frequency matrices for each simulation reveal profound microstructural degeneracies, demonstrating that traditional ensemble descriptors, such as the Harwood run number, are fundamentally incapable of describing sequence specifics. To accelerate optimization, we establish a digital twin strategy utilizing an XGBoost regressor to instantaneously predict localized chain-end propagation probabilities based on target sequence motifs. Finally, we demonstrate a closed-loop inverse design strategy that passes the 2 machine learning predictions to a deterministic root-finding optimizer, isolating exact initial monomer feeds and reactivity ratios needed to create them. By training an ML surrogate using explicit chain sequences from SSA, this framework provides a powerful alternative to classical macrokinetic equations, establishing an immediate, predictive tool to accelerate the discovery and reliable synthesis of sequence-controlled functional macromolecules.
Related
- Beyond conventional sequence-based encoding: a generalized Machine Learning pipeline for oligonucleotide property prediction
- Machine learning-accelerated evolutionary Monte Carlo for rapid phase exploration of compositionally complex materials in reactive environments
- Harnessing Machine Learning for Accelerated Drug Discovery: Opportunities and Unmet Challenges
- Controlled image-based prediction of soil nutrient concentrations (NPK) using machine and deep learning
- Machine-Learned Dynamical Representations for Accelerated RiteWeight Convergence
BibTeX
@misc{foster2026machine,
title = {{Machine Learning-Accelerated Inverse Design of Sequence-Controlled Terpolymers}},
author = {Jeffrey C. Foster},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15009833/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009833/v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - Machine Learning-Accelerated Inverse Design of Sequence-Controlled Terpolymers
AU - Jeffrey C. Foster
PY - 2026
DA - 2026/10/04
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009833/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009833/v1
AB - Sequence-controlled polymers hold immense potential for advanced materials design, yet discovering their experimental recipes is bottlenecked by the high-dimension kinetic parameter space of multicomponent copolymerization. Traditional ordinary differential equation (ODE) solvers provide rapid macrokinetic predictions but are fundamentally blind to explicit backbone microstructures. To bridge this gap, we use a stochastic simulation algorithm (SSA) engine to generate high-throughput terpolymer sequence data from a physically constrained parametric sweep across a 9-dimensional kinetic space, charting how feed stoichiometry and kinetic propensities dictate macromolecular structure. Information-theoretic analysis and t-SNE dimensionality reduction of the resulting 27-component triad frequency matrices for each simulation reveal profound microstructural degeneracies, demonstrating that traditional ensemble descriptors, such as the Harwood run number, are fundamentally incapable of describing sequence specifics. To accelerate optimization, we establish a digital twin strategy utilizing an XGBoost regressor to instantaneously predict localized chain-end propagation probabilities based on target sequence motifs. Finally, we demonstrate a closed-loop inverse design strategy that passes the 2 machine learning predictions to a deterministic root-finding optimizer, isolating exact initial monomer feeds and reactivity ratios needed to create them. By training an ML surrogate using explicit chain sequences from SSA, this framework provides a powerful alternative to classical macrokinetic equations, establishing an immediate, predictive tool to accelerate the discovery and reliable synthesis of sequence-controlled functional macromolecules.
ER -