Machine learning-assisted investigations toward polymer synthesis

Abstract

Polymers are ubiquitous in human life, with applications ranging from commodity plastics to high-tech products. The synthesis of novel structured polymers lays the foundation for developing high-performance polymeric materials crucial for meeting great interest in various fields across aerospace, biomedicine, and electronics, among many others. However, the vast chemical space of polymer structures, coupled with the complexities inherent in polymerization progress, which makes it nearly impossible to explore the entire space via conventional trial-and-error method, poses substantial challenges in predicting the polymer performances with varied structures and in tailoring polymer structures and polymerization conditions as to achieve desired properties (e.g., mechanical properties, dielectric performance, and biological functions). Recently, machine Learning (ML) techniques have emerged as illuminating avenues toward revolutionizing polymer chemistry, demonstrating significant potential in effectively navigating through the enormous landscape. Data-driven approaches could not only unveil intricate relationships among variables (e.g., polymer structures, reaction conditions, and polymer properties) but also offer unintuitive mechanistic insights that enable a deeper understanding of the polymerization process, which was previously unavailable via traditional analytical methods. To actually realize the aforementioned ideal visions, chemists employ outcomes from experiments and/or simulations as the initial database, whose quantity and quality largely determine the accuracy and reliability of subsequent modeling with delicately selected ML algorithms. Then, ML models are capable of efficiently mapping the linkage between conditions, structures, and properties from inputted data, facilitating property prediction with unprecedented efficiency and providing systematical guides for polymer structure or condition parameter optimization. Furthermore, researchers believe that transforming conventional manual operations into autonomous synthesis platforms represents an important leap toward intelligent synthesis, where automated synthesis machines with ML analysis ability iteratively generate high-quality data on a large scale, drastically shortening material development timelines. To date, polymer science integrated with ML has witnessed transformative advancements in several key fields: (1) ML-assisted structure-property relationship prediction could achieve ultra-fast yet accurate prediction toward polymer properties, thus avoiding the necessity for extensive experimental validations, which are typically fraught with labor-intensive and resource-consuming endeavors; (2) ML models are equipped to reverse the prediction from structures or conditions to properties, especially while leveraging advanced ML techniques, such as genetic algorithms and Bayesian optimization, which are able to achieve the inverse design of innovative polymeric material with targeted properties by exploring the unknown regions of chemical spaces beyond inputted datasets, heralding a paradigm shift in polymer chemistry; (3) the integration of artificial intelligence (AI) into bio-macromolecular research has exerted a significant influence on protein science and holds promise for promoting research in nucleic acids and polysaccharides, which facilitates novel discoveries in biological materials with tailored functionalities. Despite these impressive advancements, ML-driven polymer synthesis encounters several limitations to be addressed: (1) The stochastic nature of synthetic polymers, distinct from proteins and DNA, hinders the development of machine-readable structural representations of polymers, whose improved programmability is commonly regarded as an essential prerequisite for employing ML in polymer synthesis; (2) data scarcity, which is common in the early stage of exploring a new synthesis system, severely affects the accuracy of ML prediction models; (3) concerns about cost-effectiveness and scalability of automated synthesizers rise along with the prospects of implementing AI into existing experimental workflows. This review introduces the latest progress in ML-assisted polymer synthesis, probes into the key challenges, and proposes potential solutions for overcoming these obstacles. It is expected to inspire further reflection on the burgeoning field, fostering the development of next-generation polymeric materials.