Designing new alloy materials with two or more elements is a complicated process that requires combining information from multiple scales—like zooming in and out between atomic structures and large-scale material properties. Traditionally, this involves experts carefully gathering knowledge, running computer simulations, performing experiments, and analyzing results, which can be slow and challenging.

Machine learning can speed this process up by using models to predict how a material’s structure and composition affect its properties, or vice versa. However, most current models are limited—they focus on narrow goals, struggle to use knowledge from different areas, and don’t adapt well to new challenges. Most importantly, they rely on pre-existing data on which they are trained, which provides an intrinsic limitation for how such models can generalize and address new scenarios.

To solve these challenges, Alireza Ghafarollahi, a postdoc in the Massachusetts Institute of Technology (MIT) Laboratory for Atomistic and Molecular Mechanics (LAMM), and Markus Buehler, the Jerry McAfee Professor in Engineering in MIT’s departments of Civil and Environmental Engineering and of Mechanical Engineering and director of LAMM created AtomAgents.  AtomAgents is a system where multiple AI programs, each with specific skills, work together in a smart and flexible way. This platform combines the power of advanced AI, like multimodal language models, with tools for data analysis, physics simulations, and more. By working together, these AI agents can handle complex tasks like designing new metallic alloys with better properties than pure metals, by not only relying on their training data but by producing new physics insights on the fly by carrying out atomistic-level simulations and reasoning over their results. The results and framework of their new AI model is published in the Proceedings of the National Academy of Sciences (PNAS).

“AtomAgents was not only created to address the challenges and limitations of existing models, but it also serves as an intelligent system that mimics the thought process and behavior of a material scientist with an overarching goal to automate the entire materials design process,” says Ghafarollahi, lead author and postdoc in the MIT Department of Civil and Enviornmental Engineering.

The advanced generative AI models can autonomously generate new physics, visualize the results, and interpret the findings to iteratively test and refine hypotheses about materials until achieving optimal outcomes. Ghafarollahi emphasizes that this level of automation was not feasible prior to the development of large language models and multimodal multi-agent systems, and believes it’s a breakthrough in the field of materials design. 

AtomAgents comprises of a network of AI agents that dynamically interact to address complex challenges requiring the retrieval of new physical data in response to a user’s query. In addition, AtomAgents employs a suite of tools that include physics simulators capable of generating new physics by modeling atomic interactions. “With the high reasoning capabilities of these AI agents, they can decide to deploy these tools in real-time to address tasks autonomously,” says Ghafarollahi. The role of human interaction in AtomAgents is primarily for monitoring the process and intervening if new queries arise or additional information is needed. The reseachers demonstrate several computational examples in the paper showing how new physics is generated on the fly to tackle alloy design challenges.

AtomAgents can accurately predict important characteristics of alloys and suggest new ways to improve them, such as mixing metals in specific ways. This system is faster and more efficient than traditional methods. While the researchers initially developed AtomAgents for alloy design, there are no inherent limitations preventing its adaptation and expansion to other material types.

“All materials consist of atoms, and their properties are derived from atomic interactions.  Therefore, AtomAgents possesses the capability to automate the design process for various materials, from polymers to batteries and more,” says Ghafarollahi.

Ghafarollahi also adds that with recent advancements in additive manufacturing of materials, it’s possible to create hundreds of material samples within a reasonable timeframe.

“AtomAgents can be expanded to process these samples, optimizing their composition or processing conditions to achieve desired results. Combined with simulation outcomes, this approach offers a robust method for designing new materials, seamlessly integrating empirical data with theoretical models to enhance the accuracy and applicability of our predictions.”