Machine learning approach discovers crystallizable organic semiconductors

Editors' notes

This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility:

fact-checked

peer-reviewed publication

trusted source

proofread

Organic semiconductors represent a transformative technology that bridges traditional electronics with the versatility of organic materials. They make flexible, wearable devices and next-generation displays possible.

Crystallizable organic semiconductors (COS) represent a subset of organic electronic materials that have garnered substantial attention in recent years due to their unique properties and potential applications. Certain COS can form well-ordered crystalline structures, which are crucial for enhancing charge transport properties. In COS, molecules are arranged in a regular, periodic pattern, facilitating more efficient charge carrier movement through the material.

In collaboration with scientists at Princeton University, Carnegie Mellon University researchers Filipp Gusev and Olexandr Isayev, the Carl and Amy Jones Professor of Interdisciplinary Science, devised a way to use machine learning (ML) to rapidly identify potential COS materials.

"This novel approach combines machine learning modeling of thermal properties of molecules with high throughput virtual screening," said Gusev, a doctoral student in Isayev's group.

"As a result, we identified three novel crystallizable molecules in platelet form. Molecules that form large-area single crystals, known as platelets, are favorable for use in electronic devices. These three molecules build upon a pool of six that our collaborators at Princeton have seen previously to form platelets when fabricated through post-deposition annealing."

Gusev is a first co-author with Princeton University's Holly Johnson and Jordan Dull on a paper titled "Discovery of Crystallizable Organic Semiconductors with Machine Learning," published in the Journal of the American Chemical Society. Gusev presented the work at the ACS Fall 2024 meeting in Denver.

The team conducted a virtual screening of almost half a million commercially available molecules. Gusev built two ML models, one predicting melting temperature and a second predicting the enthalpy of melting. These two properties were important for narrowing down the pool of candidate COS materials by estimating a third key property: crystallization driving force.

Together with a material's melting point, the crystallization driving force can act as a metric for anticipating crystal growth morphology—or resistance to thin film crystallization—in organic semiconductors. Hence, the ML model helped to prioritize a small number of candidates that collaborators at Princeton could efficiently experimentally analyze.

The Carnegie Mellon researchers trained the predictive models to screen the candidates in multiple stages. They used Bridges-2 at Pittsburgh Supercomputing Center (PSC) with support from the NSF Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program.

"Rather than going into a lab and doing thousands of experiments to search for COS materials, we trained ML models that would guide us on what molecules to pick," Isayev said. "Essentially, you start with a haystack and end up with a few needles you were looking for."

Machine learning rapidly narrowed the pool of candidates from several hundred thousand to 44.

"From these 44 candidates, we further narrowed it to 13 molecules based on our experience in crystallization," said Johnson, a doctoral student on the team at Princeton University.

Discover the latest in science, tech, and space with over 100,000 subscribers who rely on Phys.org for daily insights. Sign up for our free newsletter and get updates on breakthroughs, innovations, and research that matter—daily or weekly.

Subscribe

The use of ML predictions and expert knowledge worked in tandem to identify platelet-forming candidate materials.

Team members from Professor Barry Rand's lab at Princeton then conducted experimental work to verify ML predictions and formulated the importance of the crystallization driving force in COS. The team selected six molecules for experimental validation based on price and commercial availability. Johnson said they used differential scanning calorimetry, or DSC, to test the candidates for their thermal properties and compared them with ML predictions.

In addition to checking the accuracy of the ML predictions, this experiment reveals to the researchers a range of temperatures for the annealing process: above the glass transition temperature but below the melting point.

"We start with DSC to get the thermal values of each material, and then we go into the fabrication and optimization phase where we're able to narrow down the conditions that work best to support crystal growth and analyze the resultant morphologies," Johnson said. "Our fabrication process is straightforward, but it requires a lot of finesse.

"You begin by depositing the material you're trying to crystallize onto a substrate under vacuum, then you abruptly anneal the samples in an inert environment. Suppose the experimental conditions, such as the film thickness, the annealing temperature, and the annealing time, are correct. In that case, there is a potential to create a film of long-range, highly ordered crystalline domains."

The Princeton collaborators conducted crystallization experiments to evaluate the six molecules over several months. Half of the molecules tested experimentally crystallized as platelets, which is the ideal morphology for future device building.

COS materials offer a platform for studying the fundamental physics of charge transport in organic systems. By examining the relationship between molecular packing, crystal structure and electronic properties, researchers can gain insight into these mechanisms to support the design of new, high-performance organic electronic materials.

Gusev said that this project is one of the directions of his graduate work. He develops methods that combine ML with experiments. "The whole idea of my research is ... to provide effective data-driven guidance for experiments across different disciplines," Gusev said.

Gusev and Isayev, along with Carnegie Mellon graduate students Evgeny Gutkin and Ben Koby and Professor of Chemistry Maria Kurnikova, were part of a collaborative team with the University of British Columbia that tied for first place on the Critical Assessment of Computational Hit-Finding Experiments (CACHE) Drug Discovery Challenge #1.

The challenge involved using novel computational methods in the search for potential targets to treat Parkinson's disease, specifically to find inhibitors for the WD40 repeat (WDR) domain of leucine-rich repeat kinase 2 (LRRK2), a common genetic cause of the disease.

"Both projects—COS materials and LRRK2 kinase inhibitors—demonstrated that well-trained ML models can predict challenging properties and reliably guide experimental work," Gusev said. "This represents an opportunity for a significant shift in decision-making authority from human experts to algorithms. These capabilities mark a crucial advancement toward self-driving laboratories like CMU Cloud Lab, illustrating the collaborative potential of machine and human intelligence."

More information: Holly M. Johnson et al, Discovery of Crystallizable Organic Semiconductors with Machine Learning, Journal of the American Chemical Society (2024). DOI: 10.1021/jacs.4c05245

Journal information: Journal of the American Chemical Society

Provided by Carnegie Mellon University