Atomic force microscopy

Machine learning the mechanical response

news background
Hailu G. Kassa

Hailu G. Kassa

University of Mons - UMONS, Belgium

View Publication
Case studyProduct: Multifrequency AFM

Summary

Challenge: Intermodulation AFM can measure 60 amplitude and phase images, in one scan at normal speed (2 min. for 256 x 256 pixels) gathering information about the surface. Each pixel represents a single point in an abstract 60-dimensional feature space.

Solution: Machine learning attempts to group or cluster the pixels in this space. The ImAFM Software Suite comes with functionality for k-means clustering. Other clustering algorithms are easily programmed using the scripting interface and the Python module scikit-learn.
Figure 1

A map showing regions of similar mechanical response on a dynamically vulcanized thermal plastic alloy (DVA). Pixels were grouped in to 5 clusters using the k-means algorithm. The x-y location of the pixel is not a feature of the data, so the method blindly reconstructs the map. Force quadrature curves are shown for the pixel closest to the centroid of each cluster, thus giving the characteristic mechanical response of the region with corresponding color on the map.

Machine learning the mechanical response

Researchers aimed to characterize the complex, nanometer-scale mechanical properties of a Dynamically Vulcanized Alloy (DVA), a blend where rubber domains are dispersed in a stiff nylon matrix. Standard Tapping Mode AFM was insufficient because the phase signal convolutes stiffness, dissipation, and adhesion, making it impossible to quantitatively distinguish the varying viscoelastic responses of the different polymer phases.

The team used the MLA-3 to implement Intermodulation AFM (ImAFM), driving the cantilever at two frequencies to measure nonlinear intermodulation products. This generated a high-dimensional dataset of 42 amplitude and phase images in a single scan. Using the MLA-3's Python scripting interface, they applied K-means clustering to this rich data, sorting pixels based on their mechanical "fingerprint" without relying on user bias.

The experiment revealed a highly inhomogeneous surface with five distinct mechanical zones, including crystalline and amorphous nylon, and rubber phases with varying cross-linking. By analyzing the force quadrature curves reconstructed from the MLA-3 data, the team could fit physical models to extract specific parameters like surface relaxation time—resolving dynamics that were invisible to standard techniques.

Back to Atomic force microscopy