Scanning probe microscopy based on reinforcement learning

Scanning probe microscopy (SPM) has become an indispensable tool for characterizing the nanoscale. The core of its working principle involves raster-scanning a sharp tip in contact with a sample of interest, while maintaining a constant tip-sample interaction. This is achieved via standard proportional-integral (PI) feedback in virtually all industrial implementations. Here we explore 1) the possibility of using machine learning to automatically find the optimal PI parameters and 2) feedback based on a Markov decision process (MDP) reinforcement learning model, free from any explicit PI feedback. PI parameters automatically selected by a neural-network-based approach achieve excellent feedback performance. The MDP model is shown to perform basic feedback, but with inferior performance than that of PI feedback, possibly due to the intrinsic unpredictability of sample topography. These proofof-concepts open up new opportunities for SPM feedback optimization under a wide range of conditions.