Polymer informatics with multi-task learning

Abstract

Summary Modern data-driven tools are transforming application-specific polymer development cycles. Surrogate models that can be trained to predict properties of polymers are becoming commonplace. Nevertheless, these models do not utilize the full breadth of the knowledge available in datasets, which are oftentimes sparse; inherent correlations between different property datasets are disregarded. Here, we demonstrate the potency of multi-task learning approaches that exploit such inherent correlations effectively. Data pertaining to 36 different properties of over 13,000 polymers are supplied to deep-learning multi-task architectures. Compared to conventional single-task learning models, the multi-task approach is accurate, efficient, scalable, and amenable to transfer learning as more data on the same or different properties become available. Moreover, these models are interpretable. Chemical rules, that explain how certain features control trends in property values, emerge from the present work, paving the way for the rational design of application specific polymers meeting desired property or performance objectives.

Bioplastic Design using Multitask Deep Neural Networks

“…Homo-and copolymer data points of T g , T m , and T d , and homopolymer data points of µ g s, E, and σ b were already utilized in previous studies. [29][30][31][32][33] The copolymer data points belonging to µ g s, E, σ y , σ b , and b , and homopolymer data points of σ y and b were collected from the PolyInfo 34 repository for this study. For consistency and uniformity, only T g and T m data points measured via differential scanning calorimetry (DSC), T d data points measured via thermogravimetric analysis (TGA), and mechanical data points recorded around room temperature (300 K) were included in the data set.…”

“…Property Predictors Multitask deep neural networks with meta learners have shown best-in-class performance in past polymer informatics studies 32,33 due to their ability to utilize inherent correlations in data that helps to overcome data sparsity. Here, we create three multiproperty predictors (one for each category in Table 1) to predict, in total, 13 polymer properties using the data set and categories profiled in Table 1 and fingerprints outlined in the Methods section.…”

Bioplastic Design using Multitask Deep Neural Networks

“…Homo-and copolymer data points of T g , T m , and T d , and homopolymer data points of µ g s, E, and σ b were already utilized in previous studies. [29][30][31][32][33] The copolymer data points belonging to µ g s, E, σ y , σ b , and ǫ b , and homopolymer data points of σ y and ǫ b were collected from the PolyInfo 34 repository for this study. For consistency and uniformity, only T g and T m data points measured via differential scanning calorimetry (DSC), T d data points measured via thermogravimetric analysis (TGA), and mechanical data points recorded around room temperature (300 K) were included in the data set.…”

Representing Polymers as Periodic Graphs with Learned Descriptors for Accurate Polymer Property Predictions

“…30 The data for the other 8 properties (atomization energy, crystallization tendency, band gap chain, band gap bulk, electron affinity, ionization energy, refractive index, dielectric constant) were previously calculated with density functional theory (DFT). 18 Altogether, this dataset consists of 15,219 total datapoints, spanning 9,935 unique homopolymer structures. For the experimentally measured properties, we exclude all measurements that correspond to compound or composite polymer samples.…”

“…Recently, a wide range of machine learning (ML) algorithms have been developed to rapidly and accurately predict polymer properties in an a priori manner. 16,17 These previous approaches typically represent polymers by their monomer repeat unit which is then converted into either i) molecular descriptors that includes the presence of molecular fragments and/or property descriptors (such as the molecular weight of the molecule) [18][19][20] or ii) a molecular graph, where atoms are represented as nodes and chemical bonds are represented as edges (Figure 1a)). 17 It should be noted that these representations were first developed for molecules before being transferred to the field of polymer informatics.…”