Organic electronics could allow companies to print electronics like paper or incorporate them into clothing to power wearable electronics—if there were only better ways to control their electronic structure.
To help address this challenge, Nick Jackson, a postdoctoral fellow in the University of Chicago’s Institute for Molecular Engineering, developed a faster way of creating molecular models by using machine learning. The models dramatically accelerate the screening of potential new organic materials for electronics, and could also be useful in other areas of materials science research.
Many believe organic electronics have the potential to revolutionize technology with their high cost-efficiency and versatility, but the current manufacturing processes used to produce these materials are sensitive, and the internal structures are extremely complex. This makes it difficult for scientists to predict the final structure and efficiency of the material based on manufacturing conditions.
Shortly after Jackson began his appointment under Juan de Pablo, the Liew Family Professor in Molecular Engineering at the University of Chicago, he had the idea to tackle such problems with machine learning. He uses this technique—a way of training a computer to learn a pattern without being explicitly programmed—to help make predictions about how the molecules will assemble.
Many materials for organic electronics are built via a technique called vapor deposition. In this process, scientists evaporate an organic molecule and allow it to slowly condense on a surface, producing a film. By manipulating certain deposition conditions, the scientists can finely tune the way the molecules pack in the film.
“It’s kind of like a game of Tetris,” said Jackson, who is a Maria Goeppert Mayer Fellow at Argonne National Laboratory. “The molecules can orient themselves in different ways, and our research aims to determine how that structure influences the electronic properties of the material.”
The packing of the molecules in the film affects the material’s charge mobility, a measure of how easily charges can move inside it. The charge mobility plays a role in the efficiency of the material as a device. In order to optimize the process, collaborating with scientist Venkatram Vishwanath of the Argonne Leadership Computing Facility, the team ran extremely detailed computer simulations of the vapor deposition process.
“We have models that simulate the behavior of all of the electrons around each molecule at nanoscopic length and time scales,” said Jackson, “but these models are computationally intensive, and therefore take a very long time to run.”
To simulate entire devices, often containing millions of molecules, scientists must develop “coarser” models. One way to make a calculation less computationally expensive is to pull back on how detailed the simulation is—in this case, modeling electrons in groups of molecules rather than individually. These coarse models can reduce computation time from hours to minutes; but the challenge is in making sure the coarse models can truly predict the physical results.
This is where the machine learning comes in. Using an artificial neural network, the machine learning algorithm learns to extrapolate from coarse to more detailed models—training itself to come to the same result using the coarse model as the detailed model.
The resulting coarse model allows the scientists to screen many, many more arrangements than before—up to two to three orders of magnitude more. Armed with these predictions, experimentalists can then test them in the laboratory and more quickly develop new materials.
Materials scientists have used machine learning before to find relationships between molecular structure and device performance, but Jackson’s approach is unique, as it aims to do this by enhancing the interaction between models of different length and time scales.
Although the targeted goal of this research is to screen vapor-deposited organic electronics, it has potential applications in many kinds of polymer research, and even fields such as protein science. “Anything where you are trying to interpolate between a fine and coarse model,” he added.
Citation: “Electronic Structure at Coarse-Grained Resolutions from Supervised Machine Learning.” Jackson et al, Science Advances, March 22, 2019. Doi: 10.1126/sciadv.aav1190
Funding: Argonne Laboratory Directed Research and Development, U.S. Department of Energy
—Article originally appeared on the Argonne National Laboratory website