Is BigSMILES the Friend of Polymer Machine Learning?

Computational methods, particularly machine learning (ML), have significantly advanced the development of innovative polymers across sectors such as aerospace, environmental science, healthcare, and green energy. Traditionally, the Simplified Molecular Input Line Entry System (SMILES) notation has been the standard for representing polymer structures within ML workflows. However, the intrinsic randomness of polymers has long hindered the effectiveness of SMILES in representation learning. Recently, the introduction of BigSMILES and its extensions has facilitated a more versatile and concise representation of polymer structures. Nevertheless, whether BigSMILES outperforms SMILES in polymer ML workflows remains an open question that warrants systematic investigation. To address this gap, we conducted comprehensive experiments to evaluate the performance of BigSMILES against SMILES, utilizing convolutional neural networks (CNNs) and large language models (LLMs) across various polymer property prediction and inverse design tasks. Here we show that in 12 distinct tasks involving both copolymer and homopolymer systems, BigSMILES-based ML workflows demonstrate performance that is comparable to, if not superior to, that of SMILES. We found that due to its more compact character representation, BigSMILES enables shorter training times compared to SMILES. Despite the use of a more concise representation, BigSMILES is capable of conveying critical chemical information and monomer connectivity (for copolymers) more accurately within the LLM framework. Our results demonstrate the potential of BigSMILES in modeling complex polymers, paving the way for sophisticated cross-scale polymer ML modeling using advanced representations such as BigSMILES. We anticipate that this work serves as a starting point for facilitating the modeling of more complex polymer systems using ML, such as copolymers and polymer composites. For instance, by employing advanced polymer representation methods, it is possible to fully account for polymer chain structures, aggregate structures, and processing factors, thereby establishing more accurate polymer-property relationships than those based on SMILES, including property prediction and polymer generation across various polymer types.