Abstract
Interface structures in complex oxides remain one of the active areas of condensed matter physics research, largely enabled by recent advances in scanning transmission electron microscopy (STEM). Yet the nature of the STEM contrast in which the structure is projected along the given direction precludes separation of possible structural models. Here, we utilize deep convolutional neural networks (DCNN) trained on simulated 4D scanning transmission electron microscopy (STEM) datasets to predict structural descriptors of interfaces. We focus on the widely studied interface between LaAlO3 and SrTiO3, using dynamical diffraction theory and leveraging high performance computing to simulate thousands of possible 4D STEM datasets to train the DCNN to learn properties of the underlying structures on which the simulations are based. We validate the DCNN on simulated data and show that it is possible (with >95% accuracy) to identify a physically rough from a chemically diffuse interface and achieve 85% accuracy in determination of buried step positions within the interface. The method shown here is general and can be applied for any inverse imaging problem where forward models are present.
Introduction
Scanning Transmission Electron Microscopy (STEM) has revolutionized our ability to visualize atomic structures. However, it suffers from an inherent "shadow puppet" problem: because the electron beam passes through the entire thickness of a sample, the resulting image is a 2D projection of a 3D object. At the interface of complex oxides—such as the widely studied $LaAlO_3$ and $SrTiO_3$—this projection creates a significant structural ambiguity. If an interface appears blurred, is it due to chemical intermixing (cation diffusion) or a physical "step" or ledge buried within the sample? Distinguishing between these two phenomena is critical for engineering the next generation of superconductors and electronic devices, yet standard imaging techniques often fail to provide a clear answer.
Methodology
To resolve this ambiguity, the researchers framed the task as an inverse problem solved through forward modeling.
Simulation at Scale
Using the $\mu$STEM simulation code on the Titan and Summit supercomputers, the team generated a massive library of synthetic "ground truth" 4D STEM datasets. These simulations, totaling approximately 350GB, accounted for varying sample thicknesses (100 Å to 300 Å), different levels of chemical diffusion, and buried physical steps at various depths.
4D STEM and DCNN Architecture
Unlike standard STEM, 4D STEM records a full 2D convergent beam electron diffraction (CBED) pattern at every pixel of a 2D scan, creating a high-dimensional data cube. To process this, the authors employed a Deep Convolutional Neural Network (DCNN) with 3D convolutional layers.
A key technical innovation was the omission of pooling layers. While pooling is standard in traditional AI to reduce data size and provide translation invariance, the authors found it destroyed the precise angular and spatial information in diffraction patterns required to identify buried structures. To ensure the model could handle real-world conditions, the simulated data was augmented with Poisson noise and small rotations.
Key Results
Classification Accuracy
The DCNN demonstrated exceptional performance in distinguishing between physical roughness and chemical diffusion. The model achieved >95% accuracy (approaching 98% after 300 epochs). This suggests the network can identify subtle symmetries and intensity shifts in the 4D data that are nearly impossible for human experts to detect by eye.
Step Localization
The network was also tasked with finding the exact position of a buried step within the sample. It achieved approximately 85% accuracy in locating the step within 10% of the sample thickness. The researchers noted that while the rough-vs-diffuse classification was relatively thickness-independent, precise localization required models trained for specific sample thicknesses, as the electron beam's signal changes significantly as it propagates deeper into the material.
Discussion and Future Work
This study demonstrates a generalizable "simulate-to-train" pipeline that can be applied to various imaging problems where reliable forward models exist. However, a primary hurdle remains the "Sim-to-Real" gap. Real-world experimental data includes lens aberrations and temporal instabilities (vibrations) that were not fully captured in the initial simulations.
Future work will focus on two paths to bridge this gap:
- High-Fidelity Scaling: Increasing the simulation library by an order of magnitude to include more experimental noise variables.
- Generative Adversarial Networks (GANs): Using Cycle-GANs to "translate" experimental data into a format that matches the simulated training data, effectively aligning the microscope's output with the DCNN's learned parameters.
Summary updated