Algorithmic Design and Implementation of an Automated Synthesis Platform for Rare and Exceptional Materials Discovery

Discovery of exceptional materials that have high-performance properties, such as perovskite materials with ideal band gaps and high stability, stands as an imperative challenge for developing sustainable electronic devices to tackle decarbonization and address the larger climate crisis. However, these exceptional materials are rare – existing only in narrow hypervolumes of vast and high-dimensional material search spaces, similar to a needle-in-a-haystack. With the rise in self-driving labs, it has become tractable to search larger volumes of these high-dimensional search spaces. However, current high-throughput self-driving labs are limited by their algorithmic performance in discovering needle-in-a-haystack exceptional materials as these labs are often guided using standard Bayesian optimization methods that inherently smooth over these regions containing exceptional materials within the search space. In this contribution, we present the design of DiSCO ([Di]scovery, [S]ynthesis, [C]haracterization, and [O]ptimization), an autonomous and high-throughput platform designed specifically to target the discovery of exceptional materials for sustainable electronic device applications. We highlight the following key contributions of DiSCO: (1) design of a zooming-based Bayesian learning with hypervolume penalization to maximize the number of exceptional materials discovered per unit time and (2) vectorization of the predicted exceptional materials for ultra-high-throughput synthesis and automated characterization within high-dimensional materials search spaces.<br/> <br/>The DiSCO platform uses a method of ultra-high-throughput inkjet gradient deposition that synthesizes up to 100 unique material compositions within 20 seconds. Each material gradient is suggested by a custom Bayesian learning algorithm, entitled ZoMBI-Hop ([Zo]oming [M]emory [B]ased [I]nitialization). ZoMBI-Hop drives the discovery of exceptional materials within DiSCO by iteratively zooming the search bounds into regions of high-reward, in turn, capturing the true rough, non-convex nature of the needle-in-a-haystack optimum, rather than smoothing it over. Furthermore, ZoMBI-Hop jumps between high-reward basins in the search space, such that once an optimal region has been discovered, then that hypervolume becomes penalized for future iterations, resulting in only new high-reward regions being searched. Using this approach, we demonstrate the targeted discovery of up to 10x more exceptional materials from 6D materials search spaces, relative to previous ZoMBI implementations, given the same number of experiments. Each optimum suggested by ZoMBI-Hop is transformed from a single point into a vector that spans the extents of the search bounds using the angle of highest predicted cumulative reward along the vector. This vectorization of a single point generates a gradient of compounds that can be tractably synthesized in an ultra-high-throughput manner using DiSCO, hence, further increasing the resolution of the search space around each predicted exceptional material. In this version of DiSCO, the implemented reward function is derived from the optimization of band gap, computed autonomously for each deposited material using computer-vision-guided hyperspectral imaging. Future implementations will contain additional reward metrics such as electrical conductivity and stability. This automated characterization of band gap takes approximately 3 minutes to compute the band gap for every 100 uniquely synthesized material compositions. Overall, with the design of this DiSCO platform, it becomes tractable to acquire voluminous experimental data of exceptional materials as well as the regions surrounding those exceptional materials, resulting in the ability to confidently down-select optimal candidates with high precision for full electronic device scaled-up and testing. Discovery of high-performance band gap perovskite materials results in progress.