Machine Learning to classify, predict structure-property relationships, and detect artifacts in AFM images

While machine learning techniques to enhance acquisition and analysis of various microscopy-based techniques is growing quickly, these techniques have been slow to catch on in the scanning probe microscopy community. While there are many reasons for this, one is that SPM suffers from slow acquisition times, making large datasets typically required for ML studies difficult to attain. Despite these challenges, ML is an important tool that can significanatly enhance analysis and interpretation of conventional AFM images. We explore its role here in 3 distinct applications: imaging classification, predicting structure-property relationships, and artifact detection.

Qualitative (phase imaging) and quantitative (peak force QNM) AFM imaging methods were used to study a series of polymer blends that varied in microstructure and bulk mechanical properties. A deep learning model based on a convolutional neural net (CNN) successfully classified the polymer blends pointing to real and meaningful differences in their microstructure. A separate regression-based CNN was built to correlate the AFM images with various bulk mechanical properties such as Young's modulus, flexural modulus, yield strength, and impact toughness. While the models were successful at predicting the Young’s modulus, flexural modulus, and yield strength, they were unsuccessful at predicting the impact toughness of the material. The success or failure of the deep learning models for this structure-property prediction provide insight into whether morphological or mechanical properties of the microstructure have a stronger influence a particular bulk mechanical properties.

Finally, two approaches were used to detect artifacts in PeakForce QNM images. The first approach used machine learning to successfully classify images that were artifact-free vs. those that had artifacts. The second approach used mathematical calculations of structural similarity to predict “good” or “bad data based on similarities (or lack thereof) between various channels in a single image.