Searching for better OLEDs, the AI way
Display is the great differentiator of all personal electronics. From swanky watches to high-end wall-straddling gaming systems, the fight is on to build displays with the sharpest black, purest blue, highest pixel density, and maximum contrast to deliver the best possible viewing experience for the consumer.
In the past decade, organic light-emitting diodes, or OLEDs, have become the darlings of the display industry with their promise of cheaper, cleaner, and more efficient lighting. OLED devices contain molecules that emit light and hence do not require backlighting as conventional LCD screens do. When used in a configuration known as active-matrix OLED, or AMOLEDs, they are a popular display technology for smartphones.
Now, a group of pioneering scientists and startup companies are using artificial intelligence (AI) to search for new OLED materials, and they are beginning to see results.
“OLEDs are ideally suited for very fast displays such as those used in gaming and watching sports. In addition, they are transparent and flexible,” says Rafael Bombarelli, an assistant professor of materials processing at the Massachusetts Institute of Technology. “This makes them perfect for niche applications such as foldable smartphones and custom brake lights that go around the car.”
Today, the OLED industry is dominated by a few large players such as United Display Corp., Samsung Display, and Sony. In the past 10 years, Samsung has made about 40 percent of all OLED displays worldwide, controlling a whopping 98 percent share of the global AMOLED market.
However, makers of OLEDs need to overcome two big challenges before the devices can become the dominant technology used for displays. First is the high price of OLEDs, and second is OLEDs’ limited lifetime.
It all comes down to the materials. OLED devices consist of multiple layers. A layer at the center produces the light while additional layers help with the transfer of photons and electrical charge carriers.
At the heart of these devices are large organic molecules with an asymmetric charge distribution, created by a metal atom at their center. For today’s OLEDs, that metal atom is iridium.
“Iridium is expensive. The technology to make these molecules is also expensive—be it solution processing or vacuum processing,” says Semion Saikin, chief science officer at Kebotix, a Boston startup company that aims to speed OLED materials discovery using AI and robotics.
It’s ironic that today’s OLEDs are expensive, because being inexpensive was part of the technology’s original great promise. For one, organic molecular compounds are low-cost materials and can be mass-produced. Second, OLEDs can be printed onto any suitable substrate by an inkjet printer or even by screen printing—processes far cheaper than those required to build LCD or plasma displays.
However, the price of OLED molecules and the cost of fabricating them flawlessly in large quantities have never matched the initial hype. In addition, the molecules have proved to be less robust than expected.
Different molecules produce different colors of light. Of these, blue has been a perennial problem to create. The energy of the bond in the organic molecule typically used to create blue is the same as the light itself, which means that the molecules degrade much faster than those that emit other colors.
According to a display industry report, blue OLEDs will reach half brightness in 14,000 hours, which comes to about five years at eight hours per day. A typical computer or cell phone screen is used a bit more than that. LCD screens, by contrast, last from 25,000 to 40,000 hours.
Kyulux Blue OLED display. Credit: Kyulux
“There is no commercially available blue phosphorescence emitter, even after decades of research. The commercialization of a blue thermally activated delayed fluorescence (TADF) emitter is expected to be challenging due to its low color purity,” says Junji Adachi, chief strategy officer for Kyulux, a Japanese OLED startup. Phosphorescence and TADF are two quantum phenomena responsible for light emission in OLEDs.
Not surprisingly, new, affordable OLED materials, particularly for blue light, that have high efficiency and last a long time are an immensely enticing target for this industry.
The situation has set up a race to discover new organic light-emitting molecules that is attracting both established industry giants and startups like Kebotix, Kyulux, and Cynora.
Bombarelli explains the allure of an OLED discovery: “It’s one of the few places where small molecule discovery can be a home run,” much like the early search for new medicinal drugs. And just as penicillin, Lipitor, or chloroquine has exerted tremendous influence on our lives, so too would a truly revolutionary light-emitting molecule.
That’s not to mention the riches such a discovery would rain down on its inventors. “The intellectual property landscape for OLEDs is highly conducive for high-margin applications. It’s probably the only place besides pharma where a single molecule can claim a high price,” says Bombarelli.
The fact that we are searching for single molecules has also allowed the introduction of a potent new ally in the exploration of OLEDs: artificial intelligence.
“There are billions of potential OLED compounds,” explains Hadi Abroshan, senior scientist at Schrödinger, a New York-based company that specializes in computational software for materials modeling and discovery.
The traditional strategy for materials discovery is to fabricate and test one molecule after another in the laboratory. It relies heavily on expertise, knowledge of chemistry, and, more often, luck. “There is nothing wrong with this approach,” says Saikin, “it’s just not scalable to a billion molecules.”
In recent years, AI-based methods have allowed us to explore vast chemical spaces by leveraging theoretical insights, quantum chemistry, chemical informatics, machine learning, industrial expertise in organic synthesis, device fabrication, and optoelectronics. This has allowed us to explore all possible OLED molecules virtually, without having to create them in the lab. The result is a systematic search for the best candidates.
In 2016, Bombarelli and his collaborators wrote one of the first papers that showed how AI can be used to screen thousands of materials. They began with a digital library of more than 1.6 million molecules, each of which was tested virtually for its electronic properties, which in turn indicate ability to emit visible light.
Normally, these properties would be calculated using a mathematical framework called density functional theory (DFT), which requires several hours of computer time per sample. Bombarelli, however, trained an AI model to predict the DFT properties of molecules. To do so, he used a computer to calculate the properties of a smaller number of molecules—some 400,000 compounds—and then fed this information into the AI model. The model, in turn, learned how to do the calculations even for molecules it had never seen. This brought down the time for virtual screening of a molecule to a fraction of a second.
“We also brought down the cost of computation from a dollar per sample to less than a fraction of a cent,” says Bombarelli.
The approach significantly broadened the scope of chemical exploration, allowing millions of molecules to be screened efficiently for the first time. A so-called synthetic accessibility score was then assigned to the molecules under scrutiny, which indicated the ease of synthesis in a laboratory. The outcome was a list of 900 molecules, each with an expected quantum efficiency greater than 22 percent, a result that would be the envy of any organic chemist. Higher efficiency implies that more light is emitted from less material.
Schrödinger recently streamlined this process by introducing what is known as an active learning model. Instead of creating one large training sample that the AI model learns from, the DFT calculations are run in smaller batches in a closed-loop learning process.
Starting with a sample size of 9,000 molecules, the DFT properties of 50 molecules were computed. An AI model predicted the properties of the rest of the pool from which the top 50 candidates were selected.
The DFT data of these top 50 are then generated using the traditional calculations, which allows the AI model to compare its predictions with the ground truth. A new set of predictions are now generated, and a fresh 50 candidates are identified. This runs in cycles until the predictions from DFT and the AI models converge. The result is a faster discovery process with a smaller number of molecules for computation and experimentation.
“The active learning method only took us 85 hours altogether to come up with the best possible OLED materials. A full DFT analysis would have taken 16 times as much time. Active learning made a paradigm shift in OLED materials design,” says Hadi.
While virtual screening is a great starting point, there is a significant amount of work before these molecules can be turned into good OLED devices. As Bombarelli puts it, “There are a lot of hops between the molecule and the device.”
Industry produces OLEDs as films on 10 × 10-foot glass slabs. Manufacturing plants require advanced equipment, tremendous resources like clean rooms, and trained staff. It’s not surprising that they cost billions of dollars to set up. Manufacturers need to be sure that the right molecule has been selected and that it meets all criteria for commercial and industrial success.
“While DFT can predict some physical properties of the molecules, they cannot predict degradation or lifetime,” says Saikin. The performance of the devices therefore requires additional parameters to be learned and several more rounds of testing.
At Kebotix, virtually screened materials are synthesized in the lab and tested quickly for molecule- and device-level performance metrics. This information is used in the selection of the next round of materials. Some starting molecules are available from a custom chemical library while others can be made from easily available feedstock material. The emphasis is on doing library generation in a cost-efficient way.
Kebotix uses retrosynthesis software that, from available reagents, helps map a synthetic route to a required molecule. One cycle of synthesis and testing involves a dozen molecules.
Kyulux is using yet another AI-assisted strategy in OLED discovery. “Most of the companies trying to utilize AI for their materials discovery depend solely on computational data, because accumulating experimental data is expensive and time-consuming. Kyulux, on the contrary, has been accumulating experimental data—molecules, thin-films, and devices—as well as computational data for the past six years,” says Adachi.
A Kebotix laboratory for automated synthesis of molecular libraries. Credit: Kebotix
Besides molecular search and discovery, the company also performs device-level optimization. Adachi says they are trying to apply various machine learning (ML) models for predicting properties of OLED devices, such as external quantum efficiency, driving voltage, lifetime, and so on. These ML models include device configurational features with all other molecular features so that the models can gradually learn optimal device configurations.
The strategy seems to be paying off. Kyulux plans to start mass production of TADF in 2023 and to expand the market to large- screen TV, AR/VR, and automotive applications after 2025. Their focus is on a trademarked quantum emission phenomena called hyperfluorescence that provides highly efficient and pure color emission with cost-competitive materials. The company expects that green and red hyperfluorescence will be commercialized in 2023, and blue hyperfluorescence the following year.
“For mass production, we are developing synthesis routes for large-scale synthesis that are different from those at the laboratory level,” Adachi says. “And we are working with outsourcing partners for mass production to ensure quality of the products.”
Altogether, it is safe to say that the OLED landscape will look vastly different in just a few years from now. From computer monitors and television screens, these molecules could also make their way into watches, smart labels, and medical sensors.
It might be then a trivial task to pair each application with the right molecule due to the abundance of OLED molecules at our disposal—thanks to AI and a bunch of bold pioneers.
Vineeth Venugopal is a science writer and metamaterials researcher who loves all things and their stories.
|Enjoy this article?
Get similar news in your inbox