Computers design next-generation OLED molecules

Organic light-emitting diodes (OLEDs) are being increasingly adopted as the display technology for mobile phones, thanks to their high contrast ratio and wide viewing angle, thinness, rapid response time, and high power efficiency. The same features make them promising for larger area displays such as televisions. In addition, they can be fabricated into flexible devices, opening a door to innovative products such as wearables, foldable electronics, and shape-fitting lighting. The affordability, efficiency, and durability of OLED materials needs further improvement for the full adoption of OLED technology. However, the search for such materials is challenging. Because OLEDs are composed of organic molecules, the design space of potential OLED materials is enormous, with billions of potentially effective materials. Given that making molecules and testing them in a laboratory is costly and time-consuming, the traditional hunt-and-try approach is ineffective.

The application of artificial intelligence and big data analytics has proven transformative in a number of fields. Tasks that before seemed accessible only to the human brain, or even solely to highly trained specialists, are now routinely performed by computers in an automated way: speech and image recognition, medical diagnosis, financial trading, domination in complex board games like Go or chess, and so forth. Materials design is no exception. By combining the fields of machine learning, quantum chemistry, chemical synthesis, and device fabrication, we have set up an accelerated OLED discovery pipeline.¹

Our computer-driven tiered discovery process (see Figure 1) efficiently screens vast areas of chemical space to discover molecules for a given application, such as an OLED. The process automatically generates millions of virtual, realizable candidate molecules and gauges their potential using theoretical simulation and machine learning. The molecules with the best predicted properties are then assessed by experimental chemists and device engineers via a Web-based tool. Only the most promising compounds are down-selected, and then synthesized and tested, making sure that experimentalists direct their effort to the most promising areas. Experimental results are then fed back into the platform to improve subsequent iterations.

Figure 1. The screening process layers machine learning, quantum chemical simulation, experts' intuition, and experiments to sieve through millions of molecules and identify the most promising ones. (Figure by Lauren A Kaye.)

Three generations of OLED technology exist based on the source of the luminescence: fluorescence, phosphorescence,² and thermally assisted delayed fluorescence (TADF).³ Most current OLED displays use highly efficient but costly phosphorescent emitters for the green and red pixels and more-durable but less-efficient fluorescent emitters for the blue pixels. Recently pioneered TADF emitters attain high efficiency at low manufacturing cost by avoiding the precious metals required for phosphorescence. We directed our platform to the discovery of new TADF emitters to leverage the vast search space yet unexplored.

After identifying nearly 200 commercially available chemical fragments in the scientific literature, we assembled in excess of two million potential donor-acceptor TADF molecules in our database. By following realistic cross-coupling chemistry rules, our combinatorial molecular generator produced synthetically accessible molecules. We conducted an empirical calibration of theoretical methods with results from existing TADF molecules and built an automated quantum chemistry workflow using density functional theory (DFT). Thousands of molecules were assessed in parallel, each taking about 40 hours of computation. Figure 2 shows the results of our search: TADF emitters must optimize two essentially opposing properties: maximize brightness (high oscillator strength, f) and efficiently recover non-emissive triplet excitons by reverse intersystem crossing (small energy difference between the lowest singlet and triplet states, ΔE_ST). In addition to mapping the fundamental limits of TADF emitters, our platform enabled us to identify promising candidates that navigate this balance.

Figure 2. Abundance of molecules as a function of predicted properties. Ideal compounds would be on the top left corner of the plot. Orange triangles correspond to theoretical leads from our work, and cyan squares for those tested in the lab. The red dots reflect the properties of thermally assisted delayed fluorescence molecules reported in the literature. f: Oscillator strength. ΔEST: Difference between the lowest singlet and triplet states.

We applied machine-learning techniques to further accelerate the theoretical screening by leveraging our growing data set. Deep neural networks trained on the DFT results accurately predicted the outcome of quantum calculations, taking miliseconds rather than hours. This enabled us to rapidly discard weak compounds and focus our simulations on the most promising ones. The lead compounds identified by the computer were subject to a collective decision procedure by experimental collaborators. The molecules arising from this consensus were synthesized and tested in optoelectronic devices. External quantum efficiencies over 20% were obtained, matching the performance of the best-known TADF molecules.

In summary, we have used a computation-first, collaborative approach to generate multiple novel OLED materials that match the performance of the best human designed ones. As next steps, we intend to expand the theoretical capability to model other physical processes that take place in OLEDs, such as host-guest interactions and bulk properties.

Harvard University's Office of Technology Development has granted a license to the software platform described here to Kyulux Inc., a Kyushu University spin-off specialized in TADF materials. Funding by the Samsung Advanced Institute of Technology is acknowledged.