Speaker
Description
While rapid exploration and optimisation of solution-processable materials in self-driving laboratories (SDLs) is advanced, adapting these approaches for inorganic materials using physical vapour deposition (PVD) presents challenges due to increased experimental complexity and higher time and energy demands for sample production. It is thus critical that the SDL’s underlying algorithms learn as much as possible about the parameter space of the new materials from as few experiments as possible.
We are developing an SDL for exploring new inorganic optoelectronic materials, with two partially conflicting aims: to produce better knowledge of the new materials, and to speed up the optimisation of their properties. Our proposed end-to-end automated workflow uses magnetron co-sputtering to deposit combinatorial thin films, a thermal post-process to convert them, and subsequent analysis by various techniques. We believe this relatively fast, flexible PVD process will facilitate rapid materials exploration that addresses the speed/knowledge generation trade-off.
Here, we focus on the use of machine learning to characterise co-sputtering processes as a function of process settings, to automatically define settings for achieving specified targets, such as deposit composition. Process data including deposition-rate feedback is used to first eliminate totally unsatisfactory process conditions (via active learning) and then build a model of the remaining four-dimensional process parameter space. Here we are using Gaussian process regression and Bayesian optimisation with active learning to update the model by querying the most informative points to be tested in the experiment. In this contribution we compare different acquisition functions including negative integrated posterior variance and Bayesian active learning by disagreement, to learn the space in as few queries as possible. A geometrical model of the sputtering flux is incorporated, to predict film compositions. We also present other aspects of the workflow including our experiences using the NOMAD database for materials-processing data.