Evaluating Polymer Representations via Quantifying Structure-Property Relationships

Abstract

Machine learning techniques are being applied in quantifying structure-property relationships for a wide variety of materials, where the properly representing materials plays key roles. Although algorithms for representation learning are extensively studied, their applications to domain-specific areas, such as polymer, are limited largely due to the lack of benchmark databases. In this work, we investigate different types of polymer representations, including Morgan Fingerprint (MF), molecular embedding (ME) and molecular graph (MG), based on a benchmark database from a subset of PolyInfo. We evaluate the quality of different polymer representations via quantifying the relationships between the representations and polymer properties, including density, melting temperature and glass transition temperature. Different representation learning schemes, such as supervised learning, semi-supervised learning and transfer learning, are investigated. It is found that ME outperforms the other representations for structure-property relationship quantification in all cases studied, and MG is shown to be much inferior than ME and MF, likely due to the relatively small volumes of training data available. For MEs, it is found that the similarities of substructure MEs under different learning schemes (e.g., SL, SSL and TL) are differently estimated, thus leading to different performance scores in structure-property relation quantification. Several ME mixtures have shown to outperform the single MEs in the corresponding regression tasks, and this is attributed to the information gain when mixing different ME.

Benchmarking Machine Learning Models for Polymer Informatics: An Example of Glass Transition Temperature

“…Their performances are generally slightly worse than the previous two groups of models. A R 2 around 0.85 and an MAE around 30 using two-monomer structure representation were reported in the literature for T g prediction, 43 which is at the same level as our single monomer-based models. Only through the evaluation on Data set_2, it is realized that the model has a relatively poor generalization ability.…”

“…Monomers, repeat units, and oligomers corresponding to several repeat units chained together are representatives of long-chain polymers. 11,16,43 Compared to monomers, the repeat unit contains bonding information indicated by the "*" symbol in their SMILES representation, which explicitly suggests the polymerization point of a polymer. Chaining several repeat units together incorporates more structure information in the originally long-chain polymer, but whether an oligomer is better than a monomer or a repeat unit remains ambiguous.…”

“…Many models claim to be able to produce good T g predictions, and some of them have been compared together. For example, Luo and co-workers 43 (1) represented polymers using a two-monomer structure (through which the bonding information between neighboring monomers was also included); (2) used different types of polymer representations like Morgan fingerprint, molecular embedding, and molecular graph; and (3) compared RF, SVM, and FFNN models in terms of T g prediction. Their best ML model was found to be the SVM model using molecular embedding, with prediction performance with R 2 = 0.865.…”

Thermal Transport in Polymers: A Review

“…Besides ML models, recent findings showed that properly representing polymer molecules can be important for constructing accurate surrogate models, since representations describe the chemistry of molecules. 249 It was found that a deep learning representation scheme (i.e., Mol2Vec) based on the Natural Language Process (NLP) algorithm was better than the conventionally used Morgan Fingerprint, a one-hot encoding scheme based on the simple chemical connections. Even if a chemistry-thermal conductivity relation can be identified from ML, properly including processing conditions as descriptors could be another obstacle for eventually validate the model prediction, as processing conditions can impact the morpology which in turn influence thermal conductivity.…”

A Machine Learning Method for Material Property Prediction: Example Polymer Compatibility

“…Researchers have made active and effective attempts to apply ML method in exploring polymer syntheses and polymer materials [22]. The copolymer synthesis and defectivity [23,24,25], mechanical properties of polymer composites [26], liquid crystal behavior of copolyether [27], thermal conductivity [28], dielectric properties [29], glass transition, melting, and degradation temperature and quantum physical and chemical properties [30,31,32,33] have been applied with machine learning and good prediction accuracy is achieved. Muramatsu et al [34] have used ML method to investigate the relationship between the phase separation structure of polymer blend and Young's modulus, and builds a predictive framework based on two-dimensional images of polymer blend as the descriptor.…”