Design of a Robotic Probe Driven by Convolutional Neural Networks Autonomously Characterizes Perovskite Photoconductivity

High-throughput deposition of functional materials has become a core component of automated experimental systems in recent literature [1, 2]. These high-throughput deposition procedures significantly accelerate the rate at which we can explore the compositions of new materials. However, as our synthesis methods continue to accelerate, relevant methods of characterization lag behind, introducing a bottleneck. Recently, there have been advancements in high-throughput and automated optical characterization methods of functional materials [3, 4], but a gap still exists in the acceleration and automation of contact-based characterization methods, limiting the accelerated characterization of materials to only certain functional properties. Hence, in this contribution, we develop the design of a four-degree-of-freedom (4DOF) robotic four-point probe that is autonomously driven by the output of a convolutional neural network (CNN) to estimate the μτ product and bulk electrical transport quality of each material from photoconductivity at a high-throughput rate of over 400 unique measurements per hour.<br/> <br/>We build upon existing methods of contact-based electrical characterization of functional materials, such as the work presented by Bash <i>et al.</i> [5]. The proposed 4DOF robotic probe in this study advances current techniques through the utilization of integrated computer vision and CNNs to predict a set of optimal contact points, or poses, to measure for each high-throughput synthesized material that capture the maximum range of measurable variance per material. Furthermore, we implement a controllable shade-minimizing LED mount on the tip of the probe for measuring photoconductivity. The combination of these advancements enables the robot to sense, plan, and act accordingly to the input provided by computer vision, ultimately achieving full autonomy in the photoconductivity characterization procedure.<br/> <br/>We design the CNN that controls the motion of the robot such that its loss function is completely spatially differentiable for every input computer vision frame. The importance of this spatially differentiable loss function is that it enables the direct geometric optimization of poses within the neurons of the CNN rather than requiring a slow iterative optimization procedure. This means that by simply minimizing the loss of the CNN, we can rapidly generate sets of optimal poses at a rate of less than one second per material. We demonstrate the performance of this proposed 4DOF robotic probe and spatially differentiable loss function on the rapid measurement of both sheet resistance and photoconductivity across over ten thousand unique contact points in less than 24 hours for Indium Tin Oxide and high-throughput synthesized formamidinium (FA) and methylammonium (MA) mixed-cation perovskite FA_1-xMA_xPbI₃ materials. Therefore, with this design, contact-based methods of characterizing high-throughput deposited materials can be accelerated and automated while maintaining an accurate representation of variance for reliable implementation into autonomous materials discovery platforms.<br/> <br/>[1] Zeng, M., et al. (2023). High-throughput printing of combinatorial materials from aerosols. <i>Nature</i>, <i>617</i>(7960), 292–298.<br/>[2] Zhao, J., et al. (2021). High-Speed Fabrication of All-Inkjet-Printed Organometallic Halide Perovskite Light-Emitting Diodes on Elastic Substrates. <i>Advanced Materials</i>, <i>33</i>(48), 2102095.<br/>[3] Siemenn, A. E., Aissi, E., et al. (2024). Using scalable computer vision to automate high-throughput semiconductor characterization. <i>Nature Communications</i>, <i>15</i>(1), 1–11.<br/>[4] Wang, T., et al. (2023). Sustainable materials acceleration platform reveals stable and efficient wide-bandgap metal halide perovskite alloys. Matter, 6(9), 2963–2986.<br/>[5] Bash, D., et al. (2021). Multi-Fidelity High-Throughput Optimization of Electrical Conductivity in P3HT-CNT Composites. <i>Advanced Functional Materials</i>, <i>31</i>(36), 2102606.