The Mendeleev–Meyer force project

Machine learning assisted quantification of graphitic surfaces exposure to defined environments

“…5 A main paradigm of the field has been to extract information from the dynamics in relation to the properties of the surfaces. 6,7 On the other hand, the AFM field expands via two main fronts. Namely, (1) improving the instrumentation so the data itself contain more information of this tip-surface interaction 8,9 and (2) improving the data processing, modeling, and overall understanding of the phenomena in order to understand and interpret the available data.…”

“…[4][5][6] Since not even such force data are obtainable directly in AFM, however, here we employ standard dynamic AFM to first acquire data and then convert them into force data. 6,7 The cantilever oscillates with free amplitude A 0 at the free resonance and it allows reconstructing the F ts vs d curves directly from 2 experimental observables, i.e., the change in amplitude A and phase U. In short, (1) monitoring A and U while approaching the surface gives us experimental Amplitude (A) and Phase (U) Distance (APD) curves;…”

“…Further details can be found elsewhere. 22,23 Since it is well known that the tip radius R strongly affects the tip-sample interaction, the effective tip radius R was monitored before and after each experiment with the use of the critical amplitude A C method; 24 R ¼ 8 nm 6 2 nm throughout as expected according to the OLYMPUS manufacturer. This method has the advantage of being in situ and fast, i.e., it takes seconds to monitor R. Only tips for which R was found to fall into this range were used.…”

“…Each set corresponds to approximately 50 data points taken for these random sets (taken with several cantilevers with only constant being the time slot). A total of 56 sets similar to those shown in Table I (see the supplementary material) were employed to generate a library to monitor the HOPG aging process, i.e., an HOPG aging library; this library is simply our final optimized predictive model produced by our ANN, 7,13 and the selection and performance of the model library are discussed below. Figure 2 shows some of the force data and an TABLE I.…”

“…Our ANN was implemented in Matlab as explained in Ref. 7. Our implementation of the ANN also considered optimizing the number of L layers and U units per layer in the network by exploiting the F 1 score criterion described above throughout.…”