IBM researchers are racing to create more sustainable PAGs, turning to AI to help create them, faster, paving the way to the era of Accelerated Discovery.
An atom here, an atom there…
The more atoms, the more complex a molecule gets — a quasi-infinity of possible molecular configurations. Which in turn means a long, expensive and tedious trial-and-error process of material discovery, where success isn’t guaranteed.
But it doesn’t have to be this way.
With artificial intelligence now greatly boosting traditional molecular design, and with quantum computing getting ready to jump in, we are entering the era of Accelerated Discovery. The era of rapidly discovering new advanced materials, vital for the manufacturing of sustainable products that could help us address a myriad of global challenges — from climate change to reducing waste to food and energy security.
It’s still early days, but IBM researchers are already applying this new AI-boosted approach to design more sustainable materials. One team has recently created new molecules dubbed photoacid generators (PAGs). With further improvements, they could help produce more environmentally friendly computing devices.
PAGs have been around since the 1980s and play a vital role in the manufacturing of computer chips. In a process called lithography, ultraviolet light creates a three-dimensional pattern in a layer of a photosensitive material — a photoresist. The photons of the light decompose the PAG inside the photoresist to produce molecules of very strong acid. These molecules catalyze the chemical reactions that create the pattern — which defines the physical structures of a computer chip, such as transistor gates or interconnect wires.
Too slow, too costly, too risky
Improvements in photoresists and lithography played an important role in the last two decades of chip development. They’ve enabled us to pack more and more transistors on continually shrinking chips, leading to ever slimmer and more powerful gadgets.
But there’s a problem.
PAGs are one of several classes of chemical compounds that have recently come under enhanced scrutiny from global environmental regulators. Researchers have been racing to create more sustainable ones — to enable a future of “green,” sustainable computing. Unfortunately, the traditional process of discovering new materials is too slow, too costly, and too risky to address this challenge in a timely and practical manner.
“Traditionally, researchers would use their own knowledge and the information they would find in published literature to design a PAG, hoping that it would have the desired properties,” says Almaden lab-based IBM researcher and electronic materials expert Dan Sanders . “Based on that initial design, they would then follow many cycles of synthesis, characterization, and testing of candidates until they managed to create a satisfactory one. It would typically take months, sometimes years — even with the help of computers to run advanced simulations.”
So his team opted for a different approach — one with the help of AI.
The use of AI in material science is not new. But even just five years ago, AI was mostly good at predicting characteristics of a material. For example, if a researcher were to input a known molecular structure, an AI would correctly predict, say, that its melting temperature is 100 degrees Celsius. However, “industrial chemists were far more interested in applying AI to rapidly design a wide variety of molecular structures beyond human creativity,” says Seiji Takeda , an IBM researcher in Tokyo.
“Just think about it — we know materials that have a billion different known molecular configurations, but there can potentially be at least 1060 times more,” he adds. “And useful materials are only a small portion of that. It is like finding a tiny diamond lost in the Sahara.”
Enter the AI-boosted Accelerated Discovery approach — the combination of advanced computing technologies to enable researchers globally to make molecular discoveries through the cloud. Recently developed at IBM Research, it’s no longer just about predicting properties of a known material — but rather about rapidly designing brand-new materials with desired properties.
Accelerated Discovery: Full steam ahead
To create the new PAGs, the Almaden team led by Sanders and fellow researcher Dmitry Zubarev first worked with experts in photoresist materials and environmental health and safety. Meticulously, they determined all the necessary performance and sustainability properties for their intended PAG. That done, they used AI, bleeding-edge computer simulation and advanced automation technologies through hybrid cloud to design and synthesize possible PAGs — much faster than was ever done before.
“Once we had outlined the properties we wanted the molecule to have, we started collecting all the data on photoacid generators out there — tucked away in patents, academic papers, pre-prints, science books and other literature,” says Sanders. That’s a daunting task for any human. So the researchers used IBM’s Deep Search AI, developed by Peter Staar ’s team at the IBM Research lab in Zurich, to compile and explore the known scientific knowledge for PAGs. They ingested 6,000 articles and patents into the AI and created a knowledge graph with 2.2M nodes and 38M edges of known materials.
They found, though, that important property data for most of the compounds that interested them was almost completely absent from the available literature. “That was a clear gap in our knowledge,” says Sanders. To close it, the researchers turned to so-called Intelligent Simulation — AI-boosted simulation led by Ed Pyzer-Knapp ’s team at IBM’s research labs in the UK. The idea was to augment the structural dataset with the necessary optical and environmental properties required to create and train an AI model.
And not any AI model — a “generative” AI model that could design a new molecule’s structure with a specific chemical property. “A generative model is an AI technology that, after being trained by a dataset, automatically designs — or generates — new objects with features similar to the original data,” says Takeda. “For example, if you train the model using a lot of images of cats and then ask the AI to generate new images of cats that are white and fluffy, that’s what the model will do. It’ll give a lot of white and fluffy cats, each one of them absolutely unique.”
Not so much interested in pictures of cats, Takeda and his team developed a generative model for molecules instead. First, they trained it with the existing PAG structure and property data, and then asked the system to design new PAG structures with lower environmental risk properties while retaining high photosensitivity. The AI obliged, and “generated about 2,000 potential PAG candidates in just five hours,” says Takeda.
That is a lot — far too many to evaluate every single one. So the researchers used IBM’s Expert-in-the-Loop technology that integrates the knowledge of human experts to enrich the AI generative model output and prioritize the most promising and actionable candidates.
That task completed, they turned to the IBM Research team in Zurich led by Teodoro Laino that was building Automated Lab technologies. They now had to solve two remaining challenges — determining the best synthetic route to make the PAGs, and to finally synthesize them in an automated robotic chemical reactor system. Laino’s team adapted their AI-based retrosynthetic tool that rapidly identifies the best way to make organic molecules — and at last created a PAG with their cloud-based automated chemical robotic reactor system, RoboRXN .
“Clearly, our Accelerated Discovery approach has greatly sped up the development of new PAGs,” says Sanders. “We are still at the very early stages, of course. But I’m sure that in the future we’ll be able to use this approach to accelerate the discovery of new materials to help us address many sustainability challenges.”
The new PAG molecules are not the only early success of the Accelerated Discovery method. Takeda’s team also used their generative model to design a new polymer membrane that absorbs carbon dioxide better than currently used membranes in carbon capture technologies. They also designed .) a new type of sugar with specific melting temperature , a collaboration with an IBM client, Nagase.
In the future, Takeda aims to expand the capabilities of his team’s AI to a broader range of material domains, including inorganic material. That could help, for instance, to create more sustainable batteries. If damaged, batteries may give off toxic gases, and the extraction of their main ingredients — typically lithium and cobalt — can lead to environmental consequences such as water pollution and depletion.
“The possibilities are endless — one can use our generative models to create new polymers, new drugs, new light-emitting material, food ingredients, ultra-low-cost biodegradable plastic bottles, flexible or even ‘paintable’ organic solar cells, you name it,” says Takeda.
“But the main point is — we’ve now shown that Deep Search, AI-enriched simulation, AI generative models, and Autonomous labs can — together with human experts of course — greatly accelerate material design and help us move closer to a sustainable society.”