[Paper Review] Potentials and challenges of polymer informatics: exploiting machine learning for polymer design

Stephen Wu, H. Yamada+3 more|arXiv (Cornell University)|Oct 15, 2020

Machine Learning in Materials Science63 references19 citations

TL;DR

This paper reviews the potential and challenges of polymer informatics, a data-driven approach leveraging machine learning to accelerate functional polymer design. It outlines four key pillars—polymer databases, structural descriptors, predictive property models, and experimental design strategies—highlighting Bayesian optimization and reinforcement learning as key methods to balance exploration and exploitation in polymer discovery with limited data.

ABSTRACT

There has been rapidly growing demand of polymeric materials coming from different aspects of modern life because of the highly diverse physical and chemical properties of polymers. Polymer informatics is an interdisciplinary research field of polymer science, computer science, information science and machine learning that serves as a platform to exploit existing polymer data for efficient design of functional polymers. Despite many potential benefits of employing a data-driven approach to polymer design, there has been notable challenges of the development of polymer informatics attributed to the complex hierarchical structures of polymers, such as the lack of open databases and unified structural representation. In this study, we review and discuss the applications of machine learning on different aspects of the polymer design process through four perspectives: polymer databases, representation (descriptor) of polymers, predictive models for polymer properties, and polymer design strategy. We hope that this paper can serve as an entry point for researchers interested in the field of polymer informatics.

Motivation & Objective

• To identify and address the major challenges hindering the development of polymer informatics, particularly due to the complex hierarchical structure of polymers.
• To review the current state of machine learning applications in polymer design across four key areas: databases, structural representation, property prediction, and design strategy.
• To advocate for open data sharing and collaborative infrastructure to enable large-scale, data-driven polymer discovery.
• To demonstrate how experimental design algorithms like Bayesian optimization and reinforcement learning can reduce the number of costly experiments in polymer development.

Proposed method

• Systematic review of existing literature on polymer informatics, focusing on four core components: polymer databases, molecular descriptors, predictive machine learning models, and experimental design frameworks.
• Evaluation of various molecular representation techniques (descriptors) that encode polymer topology, composition, and chain architecture for machine learning input.
• Application of Bayesian optimization to balance exploration and exploitation in polymer property prediction, selecting candidates with high uncertainty or high predicted utility.
• Use of reinforcement learning to frame polymer design as a sequential decision-making problem, where agents learn optimal synthesis or processing parameters through reward-based feedback.
• Analysis of case studies where machine learning reduced experimental cycles in fiber morphology, molecular weight distribution, interphase properties, and glass transition temperature control.
• Emphasis on the integration of computational simulations and experimental data to build predictive models with reduced reliance on costly trial-and-error synthesis.

Experimental results

Research questions

• RQ1 How can machine learning overcome the challenges posed by the hierarchical and complex structural nature of polymers in data-driven design?
• RQ2 What are the most effective molecular descriptors for representing polymer structures in machine learning models?
• RQ3 To what extent can predictive models based on existing data accelerate the discovery of polymers with targeted properties?
• RQ4 How can experimental design strategies such as Bayesian optimization and reinforcement learning minimize the number of required experiments in polymer development?
• RQ5 What role do open, standardized databases play in advancing polymer informatics and enabling large-scale machine learning applications?

Key findings

• Polymer informatics enables a data-driven paradigm shift in polymer design, significantly reducing reliance on time-consuming trial-and-error experimentation.
• Despite progress, major challenges remain due to inconsistent naming, lack of open databases, and the difficulty of encoding hierarchical polymer structures into machine-readable descriptors.
• Bayesian optimization and reinforcement learning have successfully reduced experimental cycles in polymer design, with demonstrated applications in tuning fiber morphology, molecular weight distribution, and glass transition temperature.
• The integration of computational simulations with machine learning allows for faster property estimation, though parameter tuning remains a bottleneck for new polymer systems.
• Open data sharing and standardized databases are essential for scaling machine learning in polymer informatics, but remain underdeveloped due to industrial data ownership and legacy inconsistencies.
• The true potential of polymer informatics lies in freeing polymer scientists from low-level optimization, enabling focus on higher-level design innovation and theoretical advancement.