Manufacturing Bits: Jan. 10

Finding new materials with inverse design
The Singapore-MIT Alliance for Research and Technology (SMART) has found a new way to perform general inverse design, a technique that can accelerate the discovery of new materials.

The concept of inverse design is simple. Let’s say you want to develop products with select materials. In a computer, you input the desired materials and the properties that you want in a system. Then, using an algorithm, the system generates a predicted solution, according to SMART, a research enterprise established by the Massachusetts Institute of Technology (MIT) in partnership with the National Research Foundation of Singapore (NRF).

This concept isn’t new and is used in several fields. It solves a major challenge in the development of new materials and compounds. Traditionally, researchers use a screening process to find new materials. In some cases, they use a materials-property database. Then, using high-performance computing (HPC), researchers can find the right formula.

This brute-force approach works, but this is also an expensive and time-consuming process. That’s where general inverse design fits in. SMART, along with the National University of Singapore (NUS) and Nanyang Technological University (NTU), have developed a new form of inverse design. For this, researchers used a new machine learning technique that uses an algorithm to identify any material that exhibits specific properties or characteristics. A subset of artificial intelligence (AI), machine learning uses advanced algorithms in systems to recognize patterns in data as well as to learn and make predictions about the information.

In the lab, researchers from SMART, NUS and NTU used a set of data to train the algorithm. Researchers used more than 50,000 compounds in a materials database. The algorithm learns and generalizes the relationships between the chemical structures and properties of these materials.

All told, inverse design, along with machine learning, can predict novel compounds or materials that possess user-targeted characteristics. It predicts materials with target formation energies, bandgaps, and thermoelectric power factors.

In the lab, researchers tested the method. In three design cases, the framework generated 142 new crystals with the desired properties, according to researchers.

“The aim of finding more effective and efficient ways to create materials or compounds with user-defined properties has long been the focus of materials science researchers. Our work demonstrates a viable solution that goes beyond specialized inverse design, allowing researchers to explore potential materials of varying composition and structure and thus enabling the creation of a much wider range of compounds. This is a pioneering example of successful general inverse design, and we hope to build on this success in further research efforts,” said Zekun Ren, a researcher from SMART.

“Materials science researchers now have an effective and comprehensive tool that allows them to discover and create new compounds and materials by simply inputting the desired characteristics,” said Tonio Buonassisi, a professor of mechanical engineering at MIT.

Added S. Isaac P. Tian, NUS graduate student, “In the next step of this journey, an important milestone will be to refine the algorithm to be able to better predict stability and manufacturability.”

Nanomaterial R&D
Northwestern University and the Toyota Research Institute (TRI) have developed a new machine learning algorithm and library to help accelerate the synthesis of new nanomaterials.

The combination of the algorithm and a large defined dataset will enable researchers to devise new structures for use in clean energy, chemical, automotive and other applications.

Researchers developed a data-generation tool. This tool, called the Megalibrary, consists of millions of nanostructures. Each nanostructure has a distinct shape, structure and composition.

The structures are positionally encoded on a 2 x 2 square centimeter chip. The structures were deposited on chips using polymer pen lithography. This is a massively parallel nanolithography tool, which deposits features on chips at a rapid pace.

In the study, researchers used a machine learning closed-loop experimental process to guide the synthesis of polyelemental nanomaterials. Researchers also used the Megalibrary. They used this data to train the model and asked it to predict compositions of four, five and six elements that would result in a certain structural feature.

In 19 predictions, the machine learning model predicted new materials correctly 18 times — an approximately 95% accuracy rate.

“Northwestern had the synthesis capabilities and the state-of-the-art characterization capabilities to determine the structures of the materials we generate,” said Chad Mirkin, a professor at Northwestern. “We worked with TRI’s AI team to create data inputs for the AI algorithms that ultimately made these predictions about materials no chemist could predict.”

“Creating this AI capability is about being able to predict the materials required for any application,” said Joseph Montoya, senior research scientist at TRI. “The more data we have, the greater predictive capability we have. When you begin to train AI, you start by localizing it on one dataset, and, as it learns, you keep adding more and more data.”

Going forward, researchers are using the approach to find catalysts critical to fueling processes in clean energy, automotive and chemical industries.