Dueling Robots—Concurrent, Robotic, AI-Driven SAXS and SANS for Industrial Coating Optimization

Virtually all materials studies involve the fusion of multiple kinds of data and multiple experimental modes, often in an iterative way and nearly universally connected by human scientific reasoning rather than unified models. This is particularly exemplified by x-ray and neutron based measurements, typically conducted at national or international user facilities on a six-month (or longer!) lead time in a highly competitive proposal landscale. For instance, a SAXS study of multicomponent nanoparticle assembly might raise questions around component identity, prompting researchers to perform contrast-matched SANS experiments. These studies would be presented alongside one another in a paper, but would seldom be truly co-refined. The nature of large user facilities means that the experiments might be months and thousands of miles apart, with iteration between the two practically impossible.<br/><br/>This talk will describe a recent first effort toward realtime, multimodal, national-user-facility based materials optimization using two NIST Autonomous Formulation Laboratory (AFL) platforms concurrently refining the same phase map at two beamlines: the I22 synchrotron SAXS and the LARMOR SANS, both at the Rutherford Appleton Laboratory site in Oxfordshire, UK. The AFL is a highly flexible, newly developed platform that can automatically prepare a specified chemical composition using pipetting, transfer it to an instrument, and trigger a measurement. Its custom-developed, open software and hardware stack allows the highly flexible intake and reduction of this measurement data, assignment of data to phase, Gaussian process regression of a phase diagram, and selection of a next measurement point based on user-specified criteria that balance specific exploration of the landscape with scientific questions of interest. This software stack has been upgraded to include classification strategies that understand the physical advantages/disadvantages of individual techniques and next point selection that similarly contemplates expected information gain. The result is that any number of AFL platforms anywhere in the world can be set up with parallel (or analogous) physical configurations, pointed using a HTTP connection to the same AI server, and collaboratively and concurrently begin refining a common model of the same phase diagram. As a first, demonstrative materials science problem, we studied the binary co-assembly of polymer and ceramic nanoparticles used in coatings (e.g. latex paints). SAXS is sensitive nearly exclusively to the high-x, high-density ceramic component while SANS is conducted in contrast-matched solvent to study only the polymeric component. Early results show a promising acceleration of the number of samples needed to identify a target phase by the use of concurrent measurement over separate grid searches.