(538e) A Machine Learning Approach to Identifying Polymorphs and the Molecular-Scale Mechanisms By Which They Interconvert in Small-Molecule Organic Semiconductors

Applications of organic electronics date as far back as 1980’s when they found use in the photo receptor material of photo copier and laser printers. Recently they have made great headway in commercial products, most notably as organic photovoltaics, wearable devices, RFID devices, biomedical devices, organic light-emitting diodes, organic field-effect transistors, and colorfully bright displays for TVs and mobile devices. This increased interest is organic electronics is primarily due to the advantages they offer in terms of unique properties such as flexibility, stretchability, and low-cost manufacturing over conventional and the primarily silicon-dominated electronics industry. Organic semiconductors can easily be chemically modified to tune their mechanical and electrical properties.¹ However, a major difficulty in deploying organic semiconductors is their predilection to pack into multiple, structurally distinct, crystal structures (polymorphism) with differing ability to transport charge. ditert-butyl[1]benzothieno[3,2-b][1]1benzothiophene (ditBu-BTBT) is an organic semiconductor which exhibits remarkably high mobilities,² ~ 7.1 cm²/Vs. A BTBT variant, diTMS-BTBT, which features bulkier side groups, is another high performing p-type organic semiconductor, but both these molecules have been shown by the Diao group at UIUC to exhibit polymorphic behavior³. Controlling polymorphism is critically important since even slight variations in Ï€-orbital overlap can lead to orders of magnitude difference in charge carrier mobility. Very few studies exist that investigate the origin and mechanism behind polymorphic transformation. Major challenges involve the low energy barriers (O(meV)), ultrafast kinetics, and structural reversibility. A deeper understanding of the process might open the door to stabilizing metastable polymorphs with different structure and charge transport characteristics. To study these polymorphs, we conducted a detailed computational study to predict the structures of stable and metastable polymorphs, and understand inter-polymorph phase transitions in BTBT systems. While experiments have only found two polymorphs, molecular simulation is capable of identifying all the thermodynamically and kinetically stabilized polymorphs.

Finding the optimal energy structures is non-intuitive and involves iteratively choosing the chemical structure, composition, or processing conditions to synthesize new materials, and improve on the design by testing their physical properties until exhaustion is achieved. For polymorphs, this “materials discovery” task requires finding the best combinations of a candidate’s six dimensions of unit cell lengths and angles and temperature that lead to the lowest energy structure. Employing an iterative Edisonian approach is wasteful in terms of synthesis time, effort, and resources. We have studied this issue by incorporating Bayesian Optimization (BO), a Machine Learning (ML) technique, into Molecular Dynamics (MD) simulations to predict the structures of thermodynamically stable and metastable polymorphs. Our novel approach is intended to uncover the relationship between structure and materials properties, in our case minimizing the total energy as a function of the design parameter, essentially our objective function (or metric for success), and uses decision theory to guide the choice of experiments during material design in such a way that we minimize the number of experiments we need to conduct to get to an optimal solution (driven by the objective function), while still retaining the fundamental physics of the problem. Combining MD and BO allowed us to predict the total energy as a function of the spatial arrangement of atoms inside the crystal, and identify the energy-minimized optimal structures obtained by running time-consuming, expensive molecular simulations for only a small fraction (~15-20 percent) of the entire set of possible candidates (over 1000 structures).

Our secondary target was to understand the kinetics behind solid state phase transitions among polymorphs; for this task, it is imperative to study the molecular-level events leading to those transformations. However, determining the pathways leading to the transformations is not a straightforward task because of the complexity involved in characterizing molecular-scale fluctuations in crystal structures. Therefore, it is essential to establish the reaction coordinate that describes how the system progresses to effect that transformation. Except for the simplest cases, the reaction coordinate involves many degrees of freedom, and it becomes a major challenge to identify which ones are important and how they are involved in the mechanism. In this study, we are using likelihood maximization (a ML technique) to systematically screen the relevant order parameters in the region in which the transformation occurs, and approximate the reaction coordinate as a function of physically relevant parameters.⁴ Next, we calculated the free energy barrier associated with the transformation along the newly found reaction coordinate. We hypothesize that this unique computational approach will provide a fundamentals-based explanation to what triggers, or effects, phase transitions in molecular crystals. The insight gained by this method can be used to design next-generation multifunctional materials with pre-chosen properties.

References:

(1) Gao, X.; Hu, Y. Development of n-type organic semiconductors for thin film transistors: a viewpoint of molecular design. J. Mater. Chem. C 2014, 2 (17), 3099–3117 DOI: 10.1039/C3TC32046D.

(2) Schweicher, G.; Lemaur, V.; Niebel, C.; Ruzié, C.; Diao, Y.; Goto, O.; Lee, W. Y.; Kim, Y.; Arlin, J. B.; Karpinska, J.; et al. Bulky end-capped [1]Benzothieno[3,2-b]benzothiophenes: Reaching high-mobility organic semiconductors by fine tuning of the crystalline solid-state Order. Adv. Mater. 2015, 27 (19), 3066–3072 DOI: 10.1002/adma.201500322.

(3) Chung, H.; Dudenko, D.; Zhang, F.; D’Avino, G.; Ruzié, C.; Richard, A.; Schweicher, G.; Cornil, J.; Beljonne, D.; Geerts, Y.; et al. Rotator side chains trigger cooperative transition for shape and function memory effect in organic semiconductors. Nat. Commun. 2018, 9 (1), 1–12 DOI: 10.1038/s41467-017-02607-9.

(4) Peters, B.; Trout, B. L. Obtaining reaction coordinates by likelihood maximization. J. Chem. Phys. 2006, 125 (5) DOI: 10.1063/1.2234477.