A Protocol to Establish Trust in Automated Synthesis Methods by Validating the Transfer of Properties Across Scales

To accelerate the discovery and optimization of new materials, rapid synthesis methods are increasingly explored, including drop casting, ink-jet printing, spraying, and microcrystal synthesis. While these methods are undoubtedly faster, they can also produce samples with undesired phases and defects, potentially compromising the quality of information gained through characterization, and hence, the materials-discovery process itself [1, 2]. To establish trust in materials produced using rapid-synthesis methods, we present a structured protocol to compare different classes of properties with higher-fidelity, lower-throughput, traditional synthesis methods such as spin coating.<br/><br/>In our study, we explore three degrees of synthesis automation and sample sizes: (1) non-automated manual spin coating of sub-micrometer-thick films with centimeter-scale [3], (2) automated OpenTrons synthesis and drop casting of microcrystals [4], and (3) automated Archerfish multi-material printing of millimeter-sized samples [5]. The transferability of said properties across these three levels of automation and size scales is quantified by benchmarking the following properties with manual spin coating: (1) crystallographic phase, (2) elemental composition, (3) surface morphology, (4) bulk transport properties, and (5) optical band gap. We demonstrate strong transferability of optical reflectance (&gt;95% cosine similarity) and band gap (&lt;0.03 eV differential) for hybrid organic-inorganic lead-halide perovskites and strong elemental correspondence (&lt;5 at.% differential) for all-inorganic halide perovskites across the least automated to most automated workflows. While crystallographic phase has shown to transfer better for certain compounds over others, surface morphology has proven challenging to transfer for most compounds and may affect outcomes of contact measurements like four-point-probe.<br/><br/>Trust in automation hinges on proving transferability across these automation scales; thus, in closing, we present preliminary work to “bridge the gap” between automated and traditional synthesis methods.<br/><br/><b>References</b><br/>[1] J. Leeman <i>et al</i>., “Challenges in high-throughput inorganic material prediction and autonomous synthesis,” <i>PRX Energy</i> <b>3</b>, 011002 (2024). https://doi.org/10.1103/PRXEnergy.3.011002<br/>[2] A.K. Cheetham and R. Seshadri, “Artificial Intelligence Driving Materials Discovery? Perspective on the Article: Scaling Deep Learning for Materials Discovery,” <i>Chemistry of Materials</i> <b>36</b>, 3490 (2024). https://doi.org/10.1021/acs.chemmater.4c00643<br/>[3] E.H. Jung <i>et al</i>. “Efficient, stable and scalable perovskite solar cells using poly(3-hexylthiophene),” <i>Nature</i> <b>567</b>, 511-515 (2019). https://www.nature.com/articles/s41586-019-1036-3<br/>[4] T. Wang <i>et al</i>., “Sustainable materials acceleration platform reveals stable and efficient wide-bandgap metal halide perovskite alloys,” <i>Matter</i> <b>6</b>, 2963-2986 (2023). https://www.sciencedirect.com/science/article/abs/pii/S2590238523003442<br/>[5] A.E. Siemenn, E. Aissi <i>et al</i>. “Using scalable computer vision to automate high-throughput semiconductor characterization,” <i>Nature Communications</i> <b>15</b>, 4654 (2024). https://www.nature.com/articles/s41467-024-48768-2 2