Breakthroughs in materials research have significantly impacted human life since the dawn of our civilization marking the well-known historical periods of stone, bronze and iron ages. Advancement and discovery of new materials is essential not only for human well being but also for a strong economy, thus marking its importance in a gamut of industries. However, the current time frame for incorporating new classes of materials into applications is long, typically about 10 to 20 years from initial research to first commercial product launch. The main reason behind this slow and expensive process is the primitive empirical and trial-and-error procedures followed to (inverse) design the optimal material to meet or exceed a prescribed set of performance criterion.
Almost all materials that relate to advanced technologies exhibit a richness of hierarchical internal structure/microstructure with salient features at multiple length scales (spanning from the atomic to the application scale or macroscale) controlling the performance characteristics of interest for a selected application. Although there is often some intuition about what these salient features might be, validated protocols do not yet exist for reliably identifying these features, and tracking their evolution during various unit processing/synthesis steps employed in the industrial manufacture of new products/devices. These fundamental connections are generally referred to as process-structure-property (PSP) linkages, and are best archived, curated, and visualized in the structure space. This is because structure evolution during processing can be represented as a distinct pathway in the structure space and each point in the “structure” space can be associated with a single value of the property of interest. Although the structure space is not easily amenable and is an extremely large, high-dimensional space, reduced order representations (e.g., Principal Component Analysis) provide a decent visualization of structure space. However, new protocols and associated workflows are needed for rigorous statistical quantification of the hierarchical material processing pathways and their linkages with properties.
New data-science enabled protocols are critically needed for establishing high value reduced-order representations of process-structure-property linkages to realize our goal of reducing time and cost of developing and deploying advanced materials into the market by half. Also, by providing a methodology to quantify process pathways, the process planners in the manufacturing industry will be able to compare, quantify, analyze and evaluate the effectiveness of a process in terms of cost, resource-utilization, efficiency in delivering the expected properties to the material, and thus choose an optimal process recipe to achieve the desired value.
Quantification of processing pathways in microstructure space is a complex and relatively unexplored problem. Complexity arises from our lack of rigorous definition of multiscale microstructure, enormity of the multidimensional microstructure space, overwhelming number of process and performance variables, thus, making it a problem of big-data. In order to tackle the complexity of the problem, our goal is to divide the problem into the following tasks:
(a) Generate microstructure realizations for a given set of processes (experimental/simulation).
(b) Represent the set of microstructures in a statistical descriptor space.
(c) Formulate the definition for process-pathways and the optimality of the process-pathways.
(d) Extracting the optimal processing pathway, for a given set of performance characteristics.
This data-analytic framework of quantification of process pathways is expected to scale to a generic set of problems with a wide range of parameters and large numbers of microstructures. While the first step is the generation of the data based on existing physics-based models, steps (b) to (d) are a problem of data analytics and big-data. Development of methodologies for machine learning is critical to achieving the tasks specified above. One important area is to develop multiscale representation of the microstructures; a recently emerging machine learning area of representation can be a good starting point to tackle this problem. For reduced-order representations, we also need to investigate nonlinear dimension reduction methods, including several manifold learning methods that can be adapted to handle the dynamic nature of our data. The process of identifying the optimal processing pathway will greatly benefit from the methodologies developed in Markov decision process and reinforcement learning. Here we will likely need to develop both continuous-time and discrete-time models. Exploring stochastic shortest-path algorithms can prove to be a good starting point. In order to enhance the communications; collaborations; and sharing of ideas, data, and tools among the diverse set of stakeholders involved in the materials research and development enterprise we are also working towards the development of an open-source software platform.
We are seeking interested, pro-active, graduate students with a basic background in scientific programming and some exposure to materials science.