Press Release

Accelerating Materials Research With AI Keeps Johns Hopkins APL Researchers Ahead of ‘Impossible’ Challenges

From the heat shield coating for NASA’s historic Parker Solar Probe to coatings for hypersonic vehicles and adhesives for underwater applications, researchers at the Johns Hopkins Applied Physics Laboratory (APL) in Laurel, Maryland have shown how materials development is central to innovation — even enabling what was previously impossible.

But researching novel materials — and in particular, characterizing their properties to ensure they’ll function as intended — is time consuming, resource intensive and expensive.

“The standard approach to materials development is, you create a batch of 50 samples or so, examine each one in detail to determine its structure and physical properties, and then conduct tests to gauge their performance,” said Morgan Trexler, who leads the Science of Extreme and Multifunctional Materials program in APL’s Research and Exploratory Development Department (REDD). “And then you tweak one variable and do it again, and again, and again, until you get the right material.”

“Current approaches to novel material design and discovery, whether they are based on theory or experiment, require significant amounts of data,” added Christopher Stiles, a senior computational materials researcher in REDD.

In an age when artificial intelligence (AI) and machine learning (ML) models can process vast quantities of data, and additive manufacturing and synthetic methods can rapidly produce samples of new materials, a new paradigm of materials characterization is required — one that can take advantage of cutting-edge software techniques and hardware platforms. APL is working on multiple applications in this space, leveraging its expertise in AI/ML, modeling and simulation, and materials science, as well as its additive manufacturing capabilities, to accelerate the characterization and analysis pipeline.

“When characterizing materials, there are frequently trade-offs between accuracy, time and throughput,” said Eddie Gienger, a senior materials researcher in REDD. “Traditionally, you’re working with only a few samples that are physically large, and it takes time to arrive at precise values as recommended in historical standards. But when you are trying to invent a new material, you typically make tiny amounts of material, and our thinking is, you can trade precision and accuracy for high throughput and collecting a lot more data that you can use to train AI/ML models.”

Clarity From Crystal Structures

A material’s crystal structure — the organization of atoms in a material — dictates many of its properties. Analyzing and interpreting a brand-new material’s crystal structure can be a time-consuming process, but APL scientists are applying AI to make the process considerably faster.

Senior computational materials researcher Nam Q. Le is leading development of an AI model that can automatically screen materials’ crystal structures. The goal is to accelerate the discovery of new materials with desirable properties, like superconductivity.

Le’s model was trained on over 30,000 simulated X-ray diffraction patterns, including known superconductors from existing databases. As a result, the model can determine whether a synthesized material is a candidate superconductor simply from the pattern of its crystal structure. With this approach, Le explained, AI screening can free up experts’ time to spend analyzing only the best candidates.

“When you have hundreds of samples to analyze in the quest to find a new material, that makes an enormous difference in the time and resources required,” said Le. “Building a model on synthetic data is inexpensive and simple compared to conducting repeated physical experiments.”

Update

Stress Testing at Scale

High-throughput techniques can also be applied to mechanical testing — assessing properties such as tensile strength, elasticity, bend resistance, fracture resistance and so on. The standard testing process for mechanical properties proceeds at a snail’s pace, allowing for about 20 samples to be tested in a day. But within the past year, scientists and engineers at APL devised a method to improve on that by several of orders of magnitude.

The prototype system combines a robotic arm that performs tensile strength tests on an array of dozens of samples in rapid succession with a load sensor that automatically and continuously records the results. This system was able to achieve a 200-times-faster throughput compared with standard approaches. The same method can be applied to a variety of mechanical properties as well.

“We focused on tensile strength for this prototype, but this approach could be extended to bending, fracture, torsion, fatigue and so on,” said Sal Nimer, a senior materials and test engineer who led the system’s development. “We envision this technique being used to autonomously measure any design-critical property, and we have other parallel efforts in the works, for example with cyclic loading to test fatigue strength.”

Another, complementary approach is small-scale mechanical testing, which researchers use to assess tensile strength and other mechanical properties. While slower than leveraging autonomy, the method enables the collection of many data points with a relatively small amount of material, making it cost-effective and efficient compared to larger-scale testing.

Acceleration Through Simulation

In some cases, AI/ML techniques can be used to circumvent mechanical testing altogether. Materials scientist Brendan Croom led the creation of a novel ML model that can predict the elastic stress response of a material based only on its porosity. And the model is fast, which allows materials to be evaluated in real time: It can simulate the stress response in seconds, whereas a conventional numerical simulation would take hours to complete.

“We know, generally speaking, that a material with a lot of porosity may not be mechanically robust, but previously we would have to perform a mechanical test to validate and quantify that relationship,” Croom said. “With this model, however, we can quickly assess how porosity will affect the elastic stress response of the material. And because the model is trained using data from high-fidelity computer simulations, we know we can trust the results.”

Taken together, the Lab’s recent work in characterizing and analyzing novel materials removes a key bottleneck that’s been holding back the pace of materials development — namely, the gap between discovery and final product.

“In all of these cases, we are able to remove the step of time-intensive characterization and replace it with high-speed, high-throughput methods that leverage the latest innovations in AI/ML, as well as novel hardware,” Gienger said. “By advancing the state of the art in materials characterization, we can accelerate the pace of materials discovery across the entire community and realize our goal of transformative innovation at a much faster pace.”