PhD Scientific Days 2024

Theoretical and Translational Medicine II.

Automatic Detection and Sorting of Atomic Force Microscopy (AFM) Force-Distance Curves

Text of the abstract

AFM force-distance curves are complex patterns generated by the dynamic probing of the specimen using a sharp tip. Depending on the interaction between the probe and the sample, the emerging patterns can be complicated and the exact generator processes and their physical origins are hard to decipher. The standard method of sorting such curves is to let a human agent do it manually. However, with the increasing number of generated curves, it is not methodically feasible to do so. Doing it automatically, however can be a challenging task too, due to the unstructured and noisy nature of the curves. Several attempts were made to create an acceptable automatic alternative to manual curve selection. Since force-distance curves are structurally heterogeneous, a possible choice can be the usage of artificial intelligence (AI) methods.

Aims:
The main drawback of using AI for such a task is the lack of available target curves for supervised learning. This limitation can be countered to some extent by fine tuning specific neural-network based architectures (Few-shot with triplet loss architecture - Waite et al, 2023). Our scientific question is about finding a way to extend the already existing models capabilities by introducing data augmentation for the models.

Methods:
We define the appropriate metrics such that these curves can be compared with each other statistically. We use general time-series augmentation steps (eg. warping). Furthermore, we experiment with the possible simulation of force-distance curves to amplify the data of the training set. Software implementation is done in Python.

Results:
Our study is currently ongoing, however several milestones were already reached. We found statistical metrics, which can be consistently used for comparing force-distance curves. We implemented a force-distance curve simulator in Python, which mimics several curve characteristics that can make it a good candidate for data augmentation.

Conclusion:
So far, by studying the above described problem, we managed to solve several sub-tasks. We still have to test whether our data augmentation methods can give a boost to the currently available models. If we manage to really boost those models significantly, our next step will be to formalize this method within a framework, such that it can be packaged as an end-to-end solution to AFM force-distance curve selection problems.