Discovery of Crystallizable Organic Semiconductors with Machine Learning

H. M. Johnson, Filipp Gusev, Jordan T. Dull, Yejoon Seo, Rodney D. Priestley, Olexandr Isayev, Barry P. RandJournal of the American Chemical Society 2024 · DOI

TL;DR

This paper leverages machine learning to predict the thermal properties (melting point and crystallization driving force) of organic semiconductors, enabling the discovery of new materials that can form large-area crystalline thin films. This approach successfully identified six candidate molecules, three of which were experimentally confirmed to crystallize into ideal platelet structures, paving the way for more efficient organic electronic devices.

What Problem Does This Paper Address?

Organic semiconductors (OSCs) are crucial for flexible electronics like OLEDs and OPVs. While crystalline OSCs offer superior performance (higher charge carrier mobility and exciton diffusion length) compared to amorphous ones, achieving specific crystalline morphologies, particularly large-area "platelets," is challenging. Current methods for inducing crystallization often require extensive experimental optimization. The paper addresses the difficulty in predicting which organic molecules will form these desirable platelet structures, aiming to accelerate the discovery of new materials for advanced organic electronics.

Methods

The researchers employed a multi-stage approach combining computational screening with experimental validation.

  1. Virtual Screening: They started with a large database of commercially available organic molecules (≈462,000) and applied a series of filters based on chemical properties (e.g., molecular weight, number of rotatable bonds, conjugated rings) to narrow down the candidates.
  2. Machine Learning (ML) Models: Two ML models, specifically Gradient Boosting Decision Trees (XGBoost), were trained to predict two key thermal properties: melting point ($T_m$) and crystallization driving force ($\Delta G_c$). These properties are known to correlate with platelet formation.
  3. ML-Driven Selection: The filtered library was then screened using the ML models to identify molecules predicted to have high $T_m$ and a negative $\Delta G_c$ (indicating a strong tendency to crystallize).
  4. Expert Filtering: From the ML-identified candidates, a smaller subset was selected based on commercial availability, price, and expert knowledge of molecular structures conducive to crystal formation.
  5. Experimental Validation: Six selected molecules were experimentally tested. Thin films were fabricated using vacuum thermal deposition and then annealed to induce crystallization. Polarized optical microscopy (POM) and X-ray diffraction (XRD) were used to characterize the resulting film morphology.

Datasets

  • Training Data: The ML models for $T_m$ and $\Delta H_m$ (enthalpy of melting, used to calculate $\Delta G_c$) were trained on datasets compiled from various sources, totaling approximately 267,000 molecules for $T_m$ and 5,000 for $\Delta H_m$.
  • Virtual Library: A database of ≈462,000 commercially available organic molecules served as the initial pool for screening.
  • Focused Library: After initial filtering, this was reduced to 7,742 molecules.
  • ML Library: Further ML screening resulted in 44 candidate molecules.
  • Expert Library: This was narrowed down to 13 molecules based on commercial availability and price.
  • Experimental Set: Six molecules were chosen for experimental validation.
  • Benchmark Data: The study also incorporated previously reported thermal properties and crystallization behaviors of 22 other organic molecules for comparison and validation of the ML model's predictive trends.

Key Insights and Ideas

  • Thermal Properties as Predictors: The study reinforces that high melting points ($T_m$) and significant crystallization driving forces ($\Delta G_c$) are strong indicators of a molecule's potential to form platelet-like crystalline structures.
  • ML for Material Discovery: Machine learning can effectively predict these crucial thermal properties, significantly accelerating the search for new crystalline organic semiconductor materials, a task that would be prohibitively time-consuming through purely experimental means.
  • Synergy of ML and Expert Knowledge: Combining ML predictions with expert chemical intuition is crucial for successful material discovery. ML narrows down the vast chemical space, while expert knowledge helps filter out less promising candidates or identify potential pitfalls (e.g., molecules prone to crystallizing upon deposition).
  • Platelet Morphology is Key: The paper emphasizes the superiority of platelet morphology for device performance due to long-range order and fewer grain boundaries, minimizing charge scattering and exciton recombination.
  • Beyond Simple Predictions: The study acknowledges that while $T_m$ and $\Delta G_c$ are good indicators, other factors like polymorphism (as seen with TBT) can influence crystallization behavior and are not fully captured by current ML models.

Key Results

  • Successful ML Prediction: The ML models achieved a mean absolute error (MAE) of 32 ± 1 K for $T_m$ and 6.7 ± 0.2 kJ/mol for $\Delta H_m$ in 5-fold cross-validation.
  • Experimental Validation Success: Out of six molecules selected by the ML-driven screening and expert review, three successfully crystallized into large-area platelets (rac-BINAP, TBT, and spiro-TAD). One crystallized as a spherulite (TPB-Cz), and two resisted crystallization (9DT and CZBDF).
  • 50% Success Rate for Platelets: This represents a 50% success rate in identifying molecules that form the ideal platelet morphology, a significant achievement for accelerating organic electronic material discovery.
  • Identification of Outliers: The study identified materials like CZBDF with unusually high melting points, suggesting limitations in the current predictive range and highlighting the need for further investigation into extreme cases.

Limitations and Future Work

  • Polymorphism and Other Factors: The paper acknowledges that factors like polymorphism (observed in TBT) can influence crystallization behavior and are not explicitly accounted for in the current ML model. Future work could explore incorporating these physical characteristics.
  • Strict Boundaries: The predicted $T_m$ and $\Delta G_c$ regions for different morphologies are not strict boundaries, and exceptions exist. Further research is needed to fully understand the reasons behind these deviations.
  • Model Accuracy: While the ML models perform well, the accuracy is limited by the heterogeneity of the $T_m$ data, decomposition upon melting, polymorphism, and amorphous forms, which are often missing from current databases. Improving data quality and incorporating more complex descriptors could enhance model performance.
  • "Self-Driving" Laboratories: The authors envision this work as a step towards "self-driving" laboratories, where ML algorithms and experimental expertise collaborate to autonomously discover and optimize materials. Future work will likely focus on expanding the scope of ML applications in materials science and integrating these predictive tools into automated experimental workflows.

Summary updated

Abstract

Crystalline organic semiconductors are known to have improved charge carrier mobility and exciton diffusion length in comparison to their amorphous counterparts. Certain organic molecular thin films can be transitioned from initially prepared amorphous layers to large-scale crystalline films via abrupt thermal annealing. Ideally, these films crystallize as platelets with long-range-ordered domains on the scale of tens to hundreds of microns. However, other organic molecular thin films may instead crystallize as spherulites or resist crystallization entirely. Organic molecules that have the capability of transforming into a platelet morphology feature both high melting point (<i>T</i><sub><i>m</i></sub>) and crystallization driving force (<i>ΔG</i><sub><i>c</i></sub>). In this work, we employed machine learning (ML) to identify candidate organic materials with the potential to crystallize into platelets by estimating the aforementioned thermal properties. Six organic molecules identified by the ML algorithm were experimentally evaluated; three crystallized as platelets, one crystallized as a spherulite, and two resisted thin film crystallization. These results demonstrate a successful application of ML in the scope of predicting thermal properties of organic molecules and reinforce the principles of <i>T</i><sub><i>m</i></sub> and <i>ΔG</i><sub><i>c</i></sub> as metrics that aid in predicting the crystallization behavior of organic thin films.
BibTeX
@article{johnson2024discovery,
  author = {H. M. Johnson and Filipp Gusev and Jordan T. Dull and Yejoon Seo and Rodney D. Priestley and Olexandr Isayev and Barry P. Rand},
  title = {{Discovery of Crystallizable Organic Semiconductors with Machine Learning}},
  journal = {Journal of the American Chemical Society},
  year = {2024},
  doi = {10.1021/jacs.4c05245},
}
Figure