afMLevel: Deep learning U-Net Models for Levelling Atomic Force Microscopy Images and Movies

Abstract

AFM experiments can now produce hundreds of frames per session and this rises to thousands in the case of video-rate high-speed AFM, but every image must still be levelled to remove tilt, bowing and line noise before it can be measured. Current automatic routines rely on user-chosen masks and parameters that work for some sample types but fail on others, so levelling often remains a manual, subjective bottleneck. Here, we present two U-Net models developed using machine learning for levelling AFM images without altering the sample topography. MLMask segments features so that they are excluded from conventional plane and line background fits, and MLBackground predicts the noise and artifact background directly, which is then subtracted. Both were trained on 2,001 experimental images from multiple instruments and sample types, using manually levelled images as ground truth. On 164 unseen images, MLMask routines brought 70-80% of images below a mean squared error of 0.1 relative to manual levelling, compared with 60-70% for the best multi-plane method, with a lower mean error (0.1-0.15 vs 0.15-0.3), and both models significantly outperformed fourteen classical routines (p < 0.01). Applied frame by frame, the models process 512x512 pixel HS-AFM movies at 4.7 frames.s-1. We provide afMLevel as an openly-available package for the use of other researchers. By removing the need to choose levelling parameters image by image, this package together with its models and paired training data, provides a foundation for automated AFM analysis pipelines.