Deep learning robotics using self-supervised spatial differentiation drive autonomous contact-based semiconductor characterization

Integrating autonomous contact-based robotic characterization into self-driving laboratories can enhance measurement quality, reliability, and throughput. While deep learning models support robust autonomy, current methods lack pixel-precision positioning and require extensive labeled data. To overcome these challenges, we propose a self-supervised convolutional neural network with a spatially differentiable loss function, incorporating shape priors to refine predictions of optimal robot contact poses for semiconductor characterization. This network improves valid pose generation by 20.0%, relative to existing models. We demonstrate our network's performance by driving a 4-degree-of-freedom robot to characterize photoconductivity at 3,025 predicted poses across a gradient of perovskite compositions, achieving throughputs over 125 measurements per hour. Spatially mapping photoconductivity onto each drop-casted film reveals regions of inhomogeneity. With this self-supervised deep learning-driven robotic system, we enable high-precision and reliable automation of contact-based characterization techniques at high throughputs, thereby allowing the measurement of previously inaccessible yet important semiconductor properties for self-driving laboratories.