The properties of any material are influenced by the processing steps used in making it. For example, if we need a material for a specific application we can simply iterate between processing and testing, adjusting our process as we go until we find the best combination of properties for our application:

This approach, used throughout most of human history, is the basis of blacksmithing. A traditional blacksmith knows that a particular processing route — a sequence of hammering, and heating and cooling steps — will produce a steel that is good for use as a sword, or a ploughshare, or what have you.

From the point of view of modern materials science, of course, we understand that the missing link here is the structure of the material. We know that structure is determined by how the material is processed, and that the structure of the material in turn determines its properties:

So an improved development scheme would involve understanding this connection between structure and properties, and then adjusting the processing to produce a structure that will give us the properties we need. This approach has two key advantages over blacksmithing. First, the insight we gain from the structure-properties connection provides clues about how to design materials with improved properties — we can guide the development cycle, rather than relying on trial-and-error. Second, because we understand the physics that determines both the development of structure during processing and the properties that arise from particular structures, we can develop forward models that describe how to go from processing to structure to properties:

because we can develop mathematical models of the structure-properties connection, we can predict the properties of a given structure computationally. This can be both faster and cheaper than making and testing real materials.

But what we really want to do is trace this path in the opposite direction — given a set of desirable properties, we what to know first what structure we need to make, and second how to make it. We refer to these as inverse problems:

In our group we’re focused on the second of these problems: Given a target microstructure, what processing steps are required to produce it? The forward problem is relatively straightforward (at least in conception) because we understand the physics and, given the processing conditions we can predict or simulate how the structure will evolve. But for the inverse problem we don’t know the processing conditions — that’s what we are trying to figure out! So we can’t just run our physics equations in reverse.

To solve this problem, we still use our computational forward models but we add advanced optimization schemes to efficiently search the vast range of possible processing routes to find those that will produce our target microstructure. In our group, we couple advanced models of microstructure development in metal alloys to sophisticated machine learning algorithms to achieve this inverse process design. For example, we are using it to design new alloys with improved resistance to shock loading. More broadly, we seek to accelerate the process by which new alloys with improved combinations of properties are developed for a broad range of applications.