Retrofitting Physical Vapour Deposition Equipment for Self-Driving Operation with a Modular Machine Learning Workflow

The concept of a Self-Driving Lab (SDL) – an experimental platform using automation and AI to autonomously explore and optimize materials – is gaining traction due to reduced barriers for AI application. However, the challenges involved vary with the type of synthesis processes targeted. Moreover, making SDLs a reality in a typical materials science lab may be an expensive prospect.<br/><br/>Our focus is on exploring and optimising new inorganic compound materials produced by physical vapour deposition (PVD), e.g. novel semiconductors like BaSn₂S₃ and BaZrS₃. This area presents special challenges for SDLs. First, the hardware must access the wide range of conditions under which inorganic materials can be grown and their properties tuned. Second, inorganic materials have multiple degrees of freedom in stoichiometry, defects, and structure, leading to a high-dimensional parameter space for the SDL to probe. Achieving the requisite capabilities with end-to-end automation and high throughput is difficult outside of dedicated (expensive/inflexible) cluster systems. Here, we present our approach for constructing an SDL for inorganic thin film materials based on retrofitting standard PVD hardware; making it potentially transferrable to many labs with relatively modest investments.<br/><br/>Our SDL hardware utilizes co-sputtering to deposit the metal elements of the subject material. We employ combinatorial deposition with QCM sensors to interpolate chemical composition in real time. The samples are then reacted with H₂S in a high-vacuum RTP furnace, allowing investigation of a temperature range up to 1000°C at pressures from 10^-6 to 10¹ Torr, with excellent control of the thermal pathway. These stages are automated by custom Python codes, while sample transfer is achieved with a 6-axis robot arm. The sample turnover time is approximately 10-20 minutes – rapid for such experiments. To avoid extending cycle time, characterization and analysis rely on non-contact imaging techniques. All process data is continuously dumped to a NOMAD Oasis.<br/><br/>To explore and optimise materials “from scratch”, our self-driving workflow comprises a series of machine learning modules that progressively narrow the search space, starting with sputtering alone and then the full two-step synthesis process. Each module uses active learning to iteratively train on accumulated experiment data and propose the next experiment, until a learning criterion is met. Different input data are used for each learning module, based on the principle of measuring "just enough" data for the task at hand. In the first module, active learning with a multilevel perceptron classifier eliminates sputtering conditions with low deposition rates (due to plasma quenching or flux attenuation). This constrains a second module using Bayesian multi-output Gaussian process regression, which learns deposition rates from each QCM sensor over the remaining parameter space. The outputs of this module, fitted to a flux-distribution model, delineate the space of sputtering processes yielding specific compositions (e.g., 1:2 Ba for BaSn₂S₃). In the third stage, the full process is utilized to produce samples, with input data for continued exploration derived from images processed through an autoencoder. This subdivides the remaining parameter space based on the visual outcome of experiments, leading to phase-diagram-like datasets where “phase regions” are first identified and only later assigned by small number of ex-situ experiments. Future workflow extensions include functional-property measurements (e.g., photoluminescence imaging) to enable optimization of functional properties in a given phase region.<br/><br/>In this presentation we will detail the equipment and workflow, including results on optimization of individual modules and the overall sequence, and early results on phase diagram derivation. The aim is to help inspire SDL implementation more widely, especially in PVD contexts.