Quantum Deep Descriptor: Physically Informed Transfer Learning from Small Molecules to Polymers

Abstract

In this study, we propose a physically informed transfer learning approach for materials informatics (MI) using a quantum deep descriptor (QDD) obtained from the quantum deep field (QDF). The QDF is a machine learning model based on density functional theory (DFT) and can be trained with a large database of molecular properties. The pre-trained QDF model can provide an effective molecular descriptor that encodes the fundamental quantum-chemical characteristics (i.e., the wave function or orbital, electron density, and energies of a molecule) learned from the large database; we refer to this descriptor as a QDD. We show that a QDD pre-trained with certain properties of small molecules can predict different properties (e.g., the band gap and dielectric constant) of polymers compared with some existing descriptors. We believe that our DFT-based, physically informed transfer learning approach will not only be useful for practical applications in MI but will also provide quantum-chemical insights into materials in the future. All codes used in this study are available at https://github.com/masashitsubaki.

Emerging Trends in Machine Learning: A Polymer Perspective

“…We refer the reader to an excellent general review on the topic . Within polymer science, transfer learning is increasingly being used. ,− For example, Li et al used it to reconstruct microstructures and generate structure–property predictions for nanocomposites. In this particular case, they use a deep convolutional neural net trained on a nonscientific corpus for their source domain .…”

“…Incorporating domain knowledge can range from conceptually simple to complex. Domain knowledge has commonly been used to select the appropriate features for a ML model or to enforce known constraints , such as transitional invariance. In both of these cases, less data is needed for the same accuracy for interpolation and, in many cases, extrapolation as the constraints and feature correllations do not need to be learned.…”

Low-Data Deep Quantum Chemical Learning for Accurate MP2 and Coupled-Cluster Correlations

“…Recently, a plethora of machine learning (ML) approaches have been developed to aim for a cheap surrogate model with the prediction accuracy of the reference data. A popular end-to-end mapping between the density and potential or density and energy has been established using atomic positions and nuclear charges, , or converged mean-field solutions, , essentially reminiscent of the exact density functional theory. Based on the atomistic locality for interatomic interactions, the atomistic ML algorithms for quantum chemistry have been shown to exhibit an astonishing ability in expressing, learning, and predicting high-dimensional data structures with complex hidden patterns for a variety of target properties if atomic descriptors are handcrafted carefully.…”

Transfer Learning Facilitates the Prediction of Polymer–Surface Adhesion Strength

“…Transfer learning (TL) can be a valuable technique to overcome the dilemma of insufficient data. − In TL, an ML model initially pretrained for a given task on a large data set of the source domain is utilized as the base to train a model for a new task by fine-tuning a small data set of the target domain. ,− Typically, TL can improve the model’s accuracy if the source and target domains are closely related. ,− , TL has achieved considerable success in speech recognition, , image recognition, , and natural language processing. , In addition, TL has also been successfully utilized in materials informatics studies − such as structural prediction of gas adsorption in MOFs, phonon properties in semiconductors, and thermal conductivity and electrochemical properties of polymers. However, these studies typically do not explore the explicit inverse design problem involved in materials design: what molecular structures, subject to reasonable constraints, are best for a given application.…”