AlloyGPT: Leveraging a language model to aid alloy discovery

Additive manufacturing of alloys has enabled the creation of machine parts that meet the complex requirements needed to optimize performance in aerospace, automotive, and energy applications. Finding the ideal mix of elements to use in these parts when there are countless possible combinations available is a complicated process that has been accelerated by computational tools and artificial intelligence.

With large language models (LLM), such as ChatGPT, evolving to better understand natural languages, researchers in the materials science and engineering department at Carnegie Mellon University have pioneered the potential to train LLM to understand a novel alloy physics language in a similar manner. Led by assistant professor Mohadeseh Taheri- Mousavi, they have developed AlloyGPT, which recognizes the relationship between composition, structure, and properties in order to generate novel designs for additively manufacturable structural alloys.

The AlloyGPT model, detailed in a recent paper published in npj Computational Materials, is unique in that it has dual functionality. It can accurately predict multiple phase structures and properties based on given alloy compositions, and conversely, it can suggest a comprehensive list of alloy compositions that meet given desired design goals.

“We have created an architecture that has learned the physics of alloys in order to design enhanced alloys that have the desired qualities for mechanical performance and manufacturability in a variety of applications,” said Taheri- Mousavi.

Taheri-Mousavi’s group, which focuses on structural alloy design, built the autoregressive model by developing a language for the physics of alloys and training this generative AI model. Rather than analyzing words, the model examines compositions and structural features in a sentence format to understand how the composition, structure, and properties are connected.

Unlike conventional iterative methods which often face challenges in finding all possible solutions, AlloyGPT can provide a comprehensive list of elemental combinations to produce the desired material properties requested. This is especially useful for designing gradient composition additively manufactured alloys in which gradual changes in material properties exist across a single part.

“It’s exciting to build a model that can solve prediction and design tasks simultaneously,” said Bo Ni, a postdoctoral researcher at Taher-Mousavi’s group. “It’s even more interesting when we demonstrate that AlloyGPT can synergize accuracy, diversity and robustness in problem solving.”

This language model has the potential to lay the groundwork for similar models and to accelerate material design for alloys manufactured by both traditional and additive manufacturing.

“Our approach will enable scientists to quickly discover alloys with new or improved properties, and will ultimately help industry partners to improve the speed and reduce the cost of their alloy design for various manufacturing processes,” said Taheri-Mousavi.