Real-Time Physics-Constrained Machine Learning in Multidimensional Atomic Force Microscopy

J.C. Agar
Drexel University,
United States

Keywords: neural networks, ai, ferroelectrics, electromechanical


The past few years have seen a tremendous increase in the dimensionality of AFM imaging. Increases in imaging speeds and modalities allow the acquisition of multimodal hyperspectral images that can image dynamic materials processes. These techniques are inherently noisy and outpace conventional downstream analysis approaches. Here, we discuss how neural networks can be applied to the automated and accelerated analysis and ultimately can be leveraged for active sampling and automated experiments. First, we demonstrate how symmetry-aware neural networks can interactively explore large, unstructured microscopy databases. Secondly, we discuss how neural networks can interpretably disentangle dynamic electromechanical responses of ferroelectric materials. We explore how custom regularization strategies can shape latent spaces to be physically interpretable; and highlight some pitfalls of applying machine learning methods in materials spectroscopy. We discuss how we can use physics-constrained neural networks for real-time fitting of materials spectroscopies. For example, we demonstrate how neural networks can accelerate the fitting of cantilever resonance obtained during band-excitation piezoresponse force switching spectroscopy of ferroelectrics. We rigorously demonstrate how batching (a stochastic averaging process) improves fit results (compared to conventional fitting techniques) on noisy data. Finally, we demonstrate a simple workflow to codesign neural networks to be deployed on field-programmable gate arrays for real-time streaming inference. We provide one practical benchmark by demonstrating streaming sub-millisecond streaming fitting of cantilever resonances. This provides a pathway toward real-time analysis and control of dynamic processes in scanning probe microscopies.