Scientists deploy AI agents to accelerate discovery of new materials

An AI-driven system automates a powerful simulation method used to discover new materials. The system can potentially reduce discovery time from months or years to just days.

The U.S. Department of Energy’s (DOE) Argonne National Laboratory has a long-running tradition of innovation in computational approaches to materials science. In 1964, Argonne scientist Aneesur Rahman published a landmark study that simulated a system of argon atoms. The study effectively launched the field of molecular dynamics, which refers to the use of computer simulation to reveal how systems of atoms interact.

More than 60 years later, Argonne continues to innovate in this realm. An Argonne research team has successfully demonstrated an artificial intelligence (AI)-driven system to automate a powerful simulation method that predicts how atoms in materials interact. Known as atomistic simulations, this method can potentially accelerate the discovery of materials for areas such as batteries, aerospace, and electronics.

“By automating these exhaustive investigations, we can potentially reduce the time requirements for discovering new materials from months or years to just days,” said Aditya Koneru, one of the study’s authors and an Argonne Scholar at the Argonne Leadership Computing Facility (ALCF).

“Our system lowers the barrier to use atomistic simulations and enables them to be much more widely adopted across the scientific community,” said Subramanian Sankaranarayanan, one of the study’s lead authors. Sankaranarayanan is an Argonne materials scientist and a professor in the Department of Mechanical and Industrial Engineering at the University of Illinois Chicago.

The study was a collaborative effort between researchers from Argonne’s Center for Nanoscale Materials (CNM), ALCF and the University of Illinois Chicago. CNM and ALCF are DOE Office of Science user facilities.

A need to make atomistic simulations easier to use

Scientists use atomistic simulations to better understand how atomic behaviors shape material properties such as strength and reactivity. The simulations can yield deep insights that inform the design and discovery of new materials for many different industries.

However, they can be challenging to use for scientists without specialized computational expertise. Researchers must manually configure, run and integrate a fragmented set of tools in a particular order. The tools serve various purposes, such as creating molecules, crystals and other structures, running simulations and analyzing results.

Materials discovery often requires exhaustive investigations of numerous materials and experimental parameters. Scientists may need to run dozens or even hundreds of atomistic simulations. This can exponentially increase time requirements.

For example, to determine the amount of strain that causes a particular material to break, a scientist may need to run numerous simulations that progressively increase strain. Automating simulations means a faster path to identifying the material’s break point.

“Manual atomistic simulations are complex, time-intensive and error-prone. For these advanced computational tools to accelerate discovery of innovative materials, they need to be much easier to use by a broader audience,” Sankaranarayanan said.

End-to-end automation

The team developed an AI framework that automates atomistic simulations from start to finish. The framework is a team of collaborating agents: AI systems that perform tasks, interpret data and make decisions with limited human intervention. The framework’s architecture was designed in collaboration with researchers at the Advanced Photon Source (APS), another DOE Office of Science user facility at Argonne.

“The multiagent AI framework streamlines the use of diverse tools to perform and analyze simulations,” said Katerina Vriza, a former CNM staff scientist at Argonne.

Using the framework is simple and straightforward. A human user enters a high-level prompt. The prompt can be a brief instruction like, ​“calculate the melting point of a gold-copper alloy.” In as little as a few minutes, the framework provides a detailed answer. The framework’s execution of the simulation is much faster and results in fewer errors compared to what a human can do.

The user’s prompt is delivered to an administrator AI agent that orchestrates the workflow and assigns tasks to various specialized AI agents. The specialists have responsibilities such as:

• Defining the arrangement of atoms in the materials to be simulated.
• Searching through scientific papers and databases to find the most suitable mathematical models to guide simulations.
• Creating input files for the simulations.
• Submitting simulation jobs for execution on a network of high performance computers.
• Performing calculations related to specific material properties.
• Analyzing simulation results.

Human-agent collaboration is an important aspect of the AI framework. In response to the human user’s initial prompt, the administrator agent may ask for additional details about the simulation request. The specialist agents may also have follow-up questions for the user as they perform their tasks.

“Our approach minimizes typical bottlenecks in atomistic simulation pipelines,” said Henry Chan, one of the study’s authors and a staff scientist at Argonne. ​“This allows researchers to significantly increase the throughput and scale.”

Promising results

The researchers tested the platform by having it run end-to-end simulations of several elements and alloys. The simulations calculated material properties such as crystalline structures, elastic behavior and vibration behavior. Simulations were executed on Carbon, a high performance computing cluster at CNM.

The agents’ calculations were remarkably close to those obtained through manual simulations performed by human experts on the research team. This outcome was a clear demonstration of the framework’s accuracy and reliability. The manual simulations were run at the National Energy Research Scientific Computing Center, a DOE Office of Science computing facility at DOE’s Lawrence Berkeley National Laboratory.

The framework is publicly available for use by the scientific community. It’s also customizable: Researchers can run the existing framework or adapt it for their own purposes, carrying out simulations for different classes of materials. And it can potentially be used to run autonomous robotic laboratory experiments.

“This multiagent framework represents a fundamental shift in the discovery pipeline,” said Uma Kornu, joint first author on the study and a research specialist at University of Illinois Chicago. ​“We are moving away from the manual orchestration of fragmented tools toward an era of autonomous, collaborative AI.”

Ultimately, the framework will allow many more researchers to navigate the vast complexities of the materials landscape with unprecedented speed and precision.

The research was supported by DOE’s Office of Science.