STM Image Analysis using Autoencoders
Peter Binev, Joshua Moorehead, Ayush Parambath, Luke Parrella, Rori Pumphrey, Miruna Savu
TL;DR
This work shows that Convolutional Autoencoders can learn compact, denoising-capable representations of STM images from multiple crystal lattices, using simulated data with realistic noise models. By comparing two CAE architectures (CAE-A and CAE-B), the study examines trade-offs between efficiency and capacity, and analyzes latent spaces via PCA. Quantitative metrics (MSE/SSIM) indicate solid reconstruction for simpler lattices and progressive challenge for more complex ones (notably FCC), while qualitative results reveal over-smoothing as complexity increases. The findings highlight both the promise of deep learning for automated STM image analysis and the need for improved latent interpretability, full-image coherence, and physics-informed constraints to enhance applicability to real-world materials science problems.
Abstract
This study explores the application of Convolutional Autoencoders (CAEs) for analyzing and reconstructing Scanning Tunneling Microscopy (STM) images of various crystalline lattice structures. We developed two distinct CAE architectures to process simulated STM images of simple cubic, body-centered cubic (BCC), face-centered cubic (FCC), and hexagonal lattices. Our models were trained on $17\times17$ pixel patches extracted from $256\times256$ simulated STM images, incorporating realistic noise characteristics. We evaluated the models' performance using Mean Squared Error (MSE) and Structural Similarity (SSIM) index, and analyzed the learned latent space representations. The results demonstrate the potential of deep learning techniques in STM image analysis, while also highlighting challenges in latent space interpretability and full image reconstruction. This work lays the foundation for future advancements in automated analysis of atomic-scale imaging data, with potential applications in materials science and nanotechnology.