Our vision is to develop new computational methods for dramatically accelerating the discovery of new materials, speeding up the discovery process by several orders of magnitude. More broadly, our Inspire project aims to show how a close cross-disciplinary collaboration involving materials scientists, materials chemists, computer scientists, and engineers can significantly advance and enrich the state-of the-art of each discipline.

In high-throughput materials discovery one searches for materials with new desirable properties by obtaining measurements on tens of thousands to millions of samples of multi-element systems. X-ray diffraction (XRD) is well-suited for rapidly collecting information on the atomic arrangements in an inorganic sample, but the data do not immediately reveal a crystal structure; indeed, it is not directly possible even to identify whether a sample comprises one, two, or more distinct crystalline phases. The data analysis and interpretation process is time-consuming and highly labor intensive. The grand challenge is therefore to develop efficient computational methods for data analysis and discovery, in order to gain insights from the deluge of materials science data, capturing and distilling structure and property relations of multicomponent materials. The problem of determining the combination of physical structures, as a function of chemical composition, is called the phase map identification problem.

The presence of significant background signal and experimental noise is one challenge in the analysis of XRD data. To address this, we developed and demonstrated the effectiveness of an approach in which materials are deposited on ultrathin amorphous membranes prepared on and supported by a crystalline silicon substrate, avoiding contributions to the signal from the substrate itself. An additional improvement to the data collection methodology, as compared with prior work, was to collect measurements as multiple detector images, allowing us to apply statistical methods to remove some noise and artifacts inherent in the experimental configuration.

At the beginning of the project, our state-of-the art analysis methodology for the phase map identification problem was based on integrating domain-specific scientific background knowledge about the physical and chemical properties of the materials into a satisfiability modulo theories (SMT) reasoning framework based on linear arithmetic and first-order logic. Though this method captured many of the physical characteristics of the problem, runtimes were long, especially in the cases of large data sets and experimental noise. Many combinatorial optimization problems admit backdoors, which are small subsets of the solution which, if known, can be used to efficiently deduce the remainder of the solution. We therefore explored a method for identifying solution backdoors for the phase map identification problem that exploits crowdsourcing and human computation. These backdoors were provided to our SMT-based solver as additional constraints.

We developed two human computing user interfaces to provide this information to our solvers. One was intended for crowdsourcing with non-expert users such as those available through Amazon Mechanical Turk (AMT). In this interface, visualizations of subsets of the data selected by the system are shown to the users, who identify diffraction peaks that are related to each other geometrically. Though the users do not need any background knowledge in the problem, they are identifying diffraction peaks that are discriminative in differentiating the parts of the signal originating from different crystal structures. In the interface intended for soliciting input from experts, the users are able to explore the full data set to identify those sets of samples in which the relationships between peaks are most evident, before identifying those patterns explicitly. These interfaces are provided through the UDiscoverIt website, along with source code, publications, background information, synthetic data generators, and data sets related to the project.

We also began to explore new approaches based on factor decomposition or nonnegative matrix factorization (NMF). In this model, the XRD signal of a particular sample are represented by a set of basis patterns, which are shared across the data set and represent the XRD pattern that would be produced by a particular pure crystalline powder, and their parameters including mixture weight and scaling. We developed a series of three methods implementing this approach, initially with a combinatorial method called CombiFD, then with more efficient gradient methods, AgileFD and interleaved agile factor decomposition (IAFD). These methods were increasingly efficient and effective at identifying solutions that are physically meaningful, and our latest solver typically runs in seconds or minutes, compared to hours for the prior methods. We developed a web-based interactive user interface for experts to explore data sets, solutions, and to run the solver called Phase-Mapper. This system and its solvers have been used by collaborating materials scientists to interpret several new materials systems, including one that led to the discovery of new solar light absorbers and the alloying-based tuning of an important property for their performance.

Last Modified: 12/13/2017
Modified by: Carla Gomes