Machine learning approach for reducing uncertainty in AFM nanomechanical measurements through selection of appropriate contact model

The force-displacement (FZ) curves obtained by atomic force microscopy (AFM) are one of the most commonly used methods for measuring the nanomechanical properties of engineering materials. However, the values of the elastic modulus obtained in this way inevitably contain a certain amount of uncertainty arising from the choice of contact model and the acquisition of the corresponding experimental parameters. Accordingly, this study proposes a supervised machine learning (SML) framework for selecting the appropriate contact model for evaluating the elastic modulus depending on the particular features of the FZ curve. In the proposed approach, the features of the FZ curve are extracted and supplied to the SML classifier, which learns these features and outputs the appropriate contact model accordingly.The classifier is implemented using five different classifiers and four different contact models. The SML classifiers are trained and tested using polyvinyl alcohol, polymethyl methacrylate, polydimethylsiloxane, and gold samples. The testing results show that the linear discriminant analysis (LDA) classifier provides the best contact model prediction quality for these materials. The practical feasibility of the proposed framework is demonstrated by processing the unseen FZ curves of Staphylococcus aureus bacteria. The LDA suggests the use of the modified Sneddon model with a testing accuracy of 96.8%. Overall, the results show that the proposed SML classifier provides a powerful tool for selecting the appropriate contact model and computing the corresponding elastic modulus with no prior knowledge required of the tip shape or any manual processing of the FZ curve.