Differentiable Modeling Enabled Autonomous Experimentation for Engineering Soft-Matter

Understanding the design rules of colloidal or soft matter systems requires us to develop phase maps that identify relations between synthesis or processing parameters and the underlying structure or property of resulting materials. This talk presents an autonomous experimentation workflow that conducts iterative on-demand synthesis and characterization of colloidal or soft-matter systems to develop a structure phase map in a closed loop. This eliminates the necessity of producing large, closely spaced samples in the design space—a bottleneck in the existing frameworks. We describe a data-driven computational model to map correlations between the synthesis or processing conditions to a relevant characterization curve and train it using an active learning framework. The data-driven model is trained in an end-to-end differentiable fashion to improve the predictive capabilities at locations it has not observed thus allowing us to use it as a proxy for sampling high-dimensional, complex, and time-consuming synthesis and characterization steps required to produce a phase map. Along with the methodological improvements for autonomous phase mapping, we provide a few case studies focusing on a novel class of colloidal gold nanoparticles synthesis using peptides and demonstrate the utility of generating phase maps with a differentiable model in extracting design rules and navigating the space of different gold nanoparticle morphologies. We would then describe a case study on colloidal gold nanoparticle retrosynthesis using the differentiable model to construct a self-driving lab alleviating the need for a user-defined Gaussian process kernel function that is hard to construct a priori.