AI unlocks new possibilities for materials design

We all know what it feels like to be tired. But you may be surprised to learn that materials also experience fatigue.

Material fatigue happens when a material is repeatedly subjected to external loads and forces that weaken or break the structural integrity. It is one of the primary causes of failure in everything from microelectronics to spacecraft and can be a costly problem. For a long time, scientists thought this damage was permanent, but a new multi-lab collaboration is aiming to change that.

Recent studies have shown that cracks so tiny they’re measured at the nanoscale (a billionth of a meter) can heal themselves. This breakthrough is already being applied in technologies like ​“self-healing” concrete driveways that fix their own cracks. Still, scientists are working to better understand these healing processes across different materials and conditions — and how to control and scale them for broader use.

To explore this frontier, researchers at the U.S. Department of Energy’s (DOE) Argonne National Laboratory have teamed up with colleagues from DOE’s Lawrence Livermore, Los Alamos and Sandia national laboratories and the University of Southern California in a project called MIRAGE — Microstructure Insights through Reliable/Interpretable AI and Guided Experiments. The project is led by Sandia.

“Understanding the fundamental mechanisms behind fatigue and self-healing — and how to control them — would be an enormous step forward,” said Jeffrey Larson, a computational mathematician at Argonne. ​“Researchers could guide experiments and induce targeted material responses, ultimately enabling the predictive design of robust, adaptive materials.”

But this is no small task. Materials go through complex combinations of processes occurring across different time periods and at different sizes. By leveraging DOE’s high performance computing systems and a type of AI called interpretable AI, MIRAGE is creating a way to accelerate the development of stronger, more adaptive metals whose properties can be predicted and improved. This approach is using AI not just to predict failures but also to explain how adaptations are made so that scientists better understand the underlying causes of materials fatigue.

The first step is discovery: identifying the basic mechanisms and their combinations that drive fatigue. This information will be compiled into a comprehensive library. Interpretable AI models will then analyze the data to detect physical patterns and uncover the root causes of structural fatigue. These models will also generate new hypotheses and experiments to test ideas and even guide materials to heal themselves.

“Interpretable AI is essential in this process,” said Todd Munson, senior computational scientist. ​“Regular AI focuses on making accurate predictions. Interpretable AI, which we’re using here, focuses on understanding how those predictions are made. Our goal is transparency: to help researchers understand the AI’s decisions, identify biases, optimize the model — and eventually develop tools to control or even reverse fatigue.”

Larson and Munson are focusing on optimization, ensuring that the AI models can efficiently simulate materials with a wide range of properties and behaviors. For example, they are designing methods to use even when some details are unknown, and they are creating surrogate models that allow researchers to reuse data more efficiently. All these methods will be incorporated into the AI’s decision-making pipeline.

Mathew Cherukara, a computational scientist and group leader of the Computational Science and Artificial Intelligence group at Argonne’s Advanced Photon Source (APS), will lead the development of an autonomous agentic AI framework to coordinate simulation and experiments. Agentic AI models can learn from their environment, adapt to conditions and make context-aware decisions with minimal human input. The APS is a DOE Office of Science user facility.

“With agentic AI, we’re building a system that actively learns from existing literature, experiments and the simulations it performs, and proposes the next step, adapting in real time,” Cherukara said. ​“This closes the loop between hypothesis and discovery in ways that would be impossible with traditional approaches. Over time, we are designing the system to learn from human feedback to become a more capable co-scientist.”

“The combination of these efforts will enable a self-improving system that we can use to generate hypotheses and design experiments — accelerating the discovery of the mechanisms behind fatigue and self-healing,” Larson said.

Ultimately, the MIRAGE team goals reach beyond just solving the problem of material fatigue. They aim to build a flexible framework that can be adapted and reused to investigate emerging or rare phenomena in complex systems.

The MIRAGE project is supported by DOE’s Office of Science through the Scientific Discovery through Advanced Computing program, co-funded by Advanced Scientific Computing Research and Basic Energy Sciences.