Leveraging Theory for Enhanced Machine Learning

Abstract

The application of machine learning to the materials domain has traditionally struggled with two major challenges: a lack of large, curated data sets and the need to understand the physics behind the machine-learning prediction. The former problem is particularly acute in the polymers domain.Here we aim to simultaneously tackle these challenges through the incorporation of scientific knowledge, thus, providing improved predictions for smaller data sets, both under interpolation and extrapolation, and a degree of explainability. We focus on imperfect theories, as they are often readily available and easier to interpret. Using a system of a polymer in different solvent qualities, we explore numerous methods for incorporating theory into machine learning using different machine-learning models, including Gaussian process regression. Ultimately, we find that encoding the functional form of the theory performs best followed by an encoding of the numeric values of the theory.

Generative BigSMILES: an extension for polymer informatics, computer simulations & ML/AI

“…Representing polymeric materials in a latent space for machine learning (ML) applications is a signicant challenge in the pursuit of automatically optimizing the in silico design of new materials. 21 An ideal representation should capture the diverse and stochastic nature of polymeric ensembles, exhibit robustness against small deviations (i.e., a small change in the latent space corresponds to a small change in the described polymer ensemble), and encode polymer architecture and chemistry in a manner that enables the prediction of chemical and physical properties. 22 To the best of our knowledge, despite notable progress in recent years, [23][24][25][26][27][28] none of the current embedding technologies have fully achieved this goal.…”

Emerging Trends in Machine Learning: A Polymer Perspective

“…More recently, Audus et al explored different methods for incorporating imperfect theory into ML models with the goal of improving interpolation, extrapolation and explainability. 130 Using the simple case study of the size of a single chain in different solvent qualities, they found that as one incorporates more knowledge all of the key metrics improved. They also found that, when the numerical values of the theory were encoded, predicting the difference between the theory and the data performed best, but that further improvement could be achieved by using the functional form of the theory.…”

“…The physics are directly included via the simple physical models, and the final expression is explainable due to the use of LASSO as depicted in Figure . More recently, Audus et al explored different methods for incorporating imperfect theory into ML models with the goal of improving interpolation, extrapolation and explainability . Using the simple case study of the size of a single chain in different solvent qualities, they found that as one incorporates more knowledge all of the key metrics improved.…”

Inverse Design of Complex Block Copolymers for Exotic Self-Assembled Structures Based on Bayesian Optimization

“…How to develop an efficient approach to realize rapid search for the target structure and to roughly determine its stable region is crucial for the inverse design of BCPs. Recently, there have been some pioneering reports about the development of approaches for “inverse design” of BCPs. − Fredrickson and co-workers have coupled bioinspired optimization algorithms (e.g., particle swarm optimization) with SCFT to explore some known structures efficiently in two-dimensional (2D) and three-dimensional (3D) spaces. , Nevertheless, the inverse design of BCPs is still in its infancy, and it needs urgently to develop more efficient methods. , …”

“…Schematic workflow of inverse design algorithm incorporating Bayesian optimization (BO) with SCFT for actively searching the phase region of a target structure. By maximizing the acquisition function a ( X ), BO probabilistically recommends a prediction for the next explorable parameter X (herein, composed of τ A , f A , and f C ), according to the uncertainty estimation of the objective function g ( X ) (eq S1 in the Supporting Information) by the Gaussian process regression (GPR) , on our collected SCFT data. FFT-3DCNN which is built on the convolutional neural network (CNN) in a 3D version is designed for automatically classifying the SCFT-converged structures, which as the input data to our CNN is preprocessed by the low-pass filter (LPF) based on fast Fourier transform (FFT).…”