Session S42: Machine Learning and Data in Polymer Physics

Autonomous robotic platform (PolyBot) for conjugated polymer processing

In the past 40 years, conjugated polymers have been widely studied and used in many novel applications, such as printable electronics, wearable electronics, and flexible energy devices, because of their unique characteristics of (opto)electronic property, mechanical softness, and low-cost manufacturing. The high throughput production of conjugated polymer thin films at low temperatures is a critical processing step in the manufacturing of devices. However, the processing of these nanometers-thick layers generally occurs under strong deformations/stressors and rapid flows, conditions under which equilibrium thermodynamics does not apply, resulting in significant structural variabilities in ways that are hard to predict. As a result, it remains a big challenge to control the assembly of these conjugated polymers towards desired multi-scale morphological orderings, which critically impact their solid-state properties (e.g., electronic property, optical property, mechanical property, stability, etc.). This talk will introduce the utilization of the self-driving lab -PolyBot (https://www.anl.gov/cnm/polybot) for autonomously controlling the assembly of conjugated polymers towards targeted structures for desired electronic properties.   

New Informatics Tools to Help Make Polymer Data Bigger

Data-driven research has the potential to change the way that we explore the world around us and solve new problems, but in polymer science we are limited by the way we handle data.  Challenges in communicating what we have made and a lack of widely-adopted standard representations for polymer data make it difficult to bring together the small, disparate data sets that have been performed, limiting our databases to a comparatively small number of curated data sets.  Taking inspiration from biology and small molecule chemistry where there has been rapid recent progress, we aim to develop new informatics tools that will enable the organization, search, sharing, and widespread use of polymer data to accelerate discovery and innovation.  First, we have developed standardized line notation representations for chemical structure of polymers, called BigSMILES.  Through canonicalization rules and extension to non-covalent chemistries, these cover a wide range of different polymer materials.  These innovations in structure representation directly enable the development of a new search language, BigSMARTS.  BigSMARTS uses graph-based search, like in small organic molecule search, and is able to address the challenge of stochastic structures in polymers by searching over the graphs of molecular generating functions rather than over the molecules themselves.  Finally, we demonstrate how these tools can be applied to accelerate data-driven research, including the development of machine learning models for block copolymers and the synthesis of chemically diverse libraries of polymers for high-throughput property characterization.   

MLExchange, bringing machine learning to the beamline

The scientific user facilities (SUF) of the Department of Energy (DOE) are some of the world's largest producers of scientific data from experiments, modeling, and simulation in the world.  The data generated span multidisciplinary sciences covering multi-faceted and complex interactions that require domain expertise to decipher intricate relationships within natural phenomena. There is tremendous potential in coordinating complex data analysis.  This can aid in building, optimizing the use of experimental facilities, increasing transparency, and widening the pool of available knowledge. As it is now, each experimentalist gathers data and scientific insight into a small part of complex physical phenomena. Restricted to limited pools of scientific data or summarized content of formal publications, intricate relationships are lost. Machine learning and artificial intelligence promise ready-to-use approaches to solve complex problems and accelerate data analysis and knowledge extraction. Driven by industry, machine learning frameworks are being developed at rapid speed. However, the applications in science and in particular the world of scientific user facilities is less evolved. The threshold to develop, train, and test machine learning models is still very steep and requires significant time commitment by beamline scientists. MLExchange is targeting this threshold with the development and deployment of easy-to-use solutions that target science performed at SUF's. We will present our first developments in the field of labeling, segmentation, and XRD.   

Improving prediction and understanding with theory aware machine learning

Machine learning as applied to polymer physics has recently shown immense progress---mostly in areas where there are existing large datasets or where datasets can be generated quickly. However, there are numerous interesting problems where the dataset sizes are too small or the need to understand the physics behind the machine learning prediction is essential. Here, we aim to tackle both of these problems by incorporating domain knowledge into machine learning models. Specifically, using a toy system of polymers in different solvent qualities, we compare several methods for incorporating theory into machine learning using a simple, imperfect but easily interpretable theory. We also explore the intersection these methods with different machine models including random forest and Gaussian process regression.   

Data driven approaches to quantifying charge transport in semiconducting systems

Charge transport in molecular solids, such as semiconducting polymers, is strongly affected by packing

and structural order over several length scales. We describe a computationally scalable methodology

using graph theory to explore the influence of molecular ordering on charge mobility. This model

accurately reproduces the analytical results for transport in nematic and isotropic systems, as well as

experimental results of the dependence of the charge carrier mobility on orientation correlation length

for polymers. This approach can be deployed both on continuum/experimental data as well as MD

simulation data. We illustrate with examples including (a) modeling how defect distribution (correlated

and uncorrelated) in semiconducting polymers can modify the mobility, and (b) quantifying the

resilience of charge transport. This work enables rapid (and computationally extensible) evaluation of

charge mobility semiconducting polymer devices. This work is collaborative work with the Risko group

(U Kentucky) and the Chabinyc group (UC Santa Barbara).