Imagine you’re an engineer who has solved an age-old challenge, but for one tiny detail: You lack the perfect material for the task. You’re not sure such a material exists and, even if it did, the realm of possible candidates could number in the thousands or millions.
Don’t despair, says Stanford’s Evan Reed, associate professor of materials science and engineering. He and two graduate students, Austin Sendek and Gowoon Cheon, have developed artificial intelligence algorithms that can help engineers find the perfect material for important applications. “These algorithms are like search engines for materials,” Reed says.
In three recent papers he, Sendek and Cheon describe how they use machine learning – computational systems that teach themselves – to analyze the physical and chemical properties of thousands of materials. The machine learning systems create algorithms that comb through these materials to find those that mathematically match the requirements for the task at hand. Experimental scientists can then use these computer-generated hunches to create and test materials under real-world conditions, much faster than past hit or miss methods.
In one of these recent papers, in the Journal of Physical Chemistry Letters, Reed explored a category of materials known as metal-organic frameworks (MOFs). MOFs are crystals that are structurally strong, porous at the nanoscale and inexpensive to make. So far they have been used to capture hydrogen and carbon gases within the pores of solid materials. But researchers have become interested in using MOFs for fuel cells and thermoelectric devices, and these new uses would require electrical conductivity. Unfortunately, very few known MOFs are electrically conductive. With tens-of-thousands of MOFs out there to explore, Reed felt there might be at least a few to fit the bill, so he set his computers to the task.
Reed screened almost 3,000 metal-organic frameworks to winnow to the six best candidates. His team then confirmed the electrical conductivity of these materials through detailed calculations that used standard computational methods, to offer experimentalists a new palette of MOF possibilities.
In a paper for the journal Chemistry of Materials, Sendek, a doctoral candidate in applied physics, described his efforts to identify safer lithium-based solid materials for solid-state lithium-ion batteries, as an alternative to commercial lithium-ion batteries with liquid electrolytes that have caused some well-documented smartphone and electric car fires.
For decades, the search for solid lithium-ion electrolyte materials proceeded mostly on a trial-and-error basis. Sendek used a machine learning-based approach to survey 12,000 possible candidates and identify 21 materials that were likely to exhibit a number of important properties for solid electrolytes, including high lithium-ion conductivity. He then spent over a year evaluating those candidates using conventional physics-based simulations, and found that 10 of the 21 did, in fact, exhibit the conductivity values in simulation that the machine learning model predicted – a significantly higher success rate than the intuition-guided guess-and-check searches of the past.
In fact, Sendek turned the search into a bit of a competition, pitting his algorithm against a team of six doctoral students who performed a similar hunt for fast lithium-ion conductors. His model proved twice as accurate and over a thousand times faster than the human team.
“I liken our approach to facial recognition software for materials,” Sendek says.
In a third AI-meets-materials paper, published in the Journal of Physical Chemistry Letters, Cheon, a doctoral candidate in applied physics, described her use of machine learning to predict the existence of two-dimensional materials – substances so thin that their thickness is measured in layers of atoms. The first 2D material to be discovered was graphene, which consists of a single atomic layer of graphite. The scientists who found it in 2004 shared the 2010 Nobel Prize in physics for revealing its amazing properties. Electronics researchers hope to use 2D materials to create atomically thin circuits and devices but, so far, only a few dozen of these compounds have been discovered.
To expand the list, Cheon used a physics-based machine learning model to study 16 million chemical compounds to predict which might be exfoliated into 2D layers. Her model added more than a thousand candidates to the field. Most have never been synthesized, providing experimentalists with a roadmap in the search for ultrathin materials, Cheon says.
This process is extraordinarily useful, Reed says, because relatively few materials have ever been measured, probed and tested, much less cataloged in a database. Machine learning can expedite the discovery process.
The process is complex but the logic is simple: Molecules and crystalline structures have physical and chemical properties; such properties can be expressed mathematically; computational systems are adept at mathematical analysis. “Even though we lack lab test data on most materials, we can still predict something about their likely properties computationally,” Reed says.