Real materials in action: Data analysis developments for real materials doing real things

Hypothesis: It is possible to separate overlapping signals from heterogeneous samples in time of flight neutron experiments by exploiting the high dimensionality of the 'experiment parameter space' and using modern algorithms

Modern materials under study for next generation technologies, such as for energy conversion and storage, environmental remediation and health, are highly complex, often heterogeneous and nanostructured. Examples are battery materials, nano-catalysts, ferroic materials such as ferroelectrics and soft ferromagnets, metal-insulator transition materials, thermoelectrics. The materials are prized for their properties allowing us to make advanced devices to solve problems in sustainable energy, medicine and so on. In real applications the materials can undergo dynamic changes that are at the heart of the property that we are trying to exploit. For example, ions move around under electrochemical potentials in batteries and catalysts temporarily undergo chemical changes during the catalysis process. We therefore seek to understand materials not just in their thermodynamically stable state, but also changes that occur as they are driven by external forces.

This means making structural and spectroscopic measurements in actual operating devices. In general, such devices are heterogeneous, made up of sandwiches of different materials in complex arrangements on micron and millimeter lengths-scales. A fundamental challenge in such experiments is to separate the signals from the different components in the device, so the signal of interest (for example, the nanoscale structure of the oxide battery cathode on cycling) can be extracted and successfully analyzed. The proposed work will develop robust, automated, data analysis tools to accomplish this task for neutron data.

The approach will exploit the high dimensionality of the neutron data, which exists in many thousands of detector pixels and thousands of time points, and to use both physics-based statistical methods and data analytic methods including machine learning to automatically recognize, cluster, and subsequently separate signals from different component signals in the measured pattern with a goal of being able to do fully quantitative studies of complex heterogeneous samples with neutrons.