# Prediction of EDS Maps from 4DSTEM Diffraction Patterns Using Convolutional Neural Networks

## TL;DR

This paper presents a CNN-based method to predict EDS spectra from 4D-STEM diffraction patterns, enabling faster, less damaging, and high-throughput chemical analysis of materials.

## Contribution

It introduces a novel machine learning framework that directly infers elemental compositions from diffraction data, reducing the need for traditional, time-consuming spectroscopy techniques.

## Key findings

- CNN accurately predicts elemental compositions, especially for elements with strong diffraction contrast.
- The model performs well in both interpolation and extrapolation scenarios.
- Predictions improve with additional structural context and are validated by visual and correlation analyses.

## Abstract

Understanding the relationship between atomic structure (order) and chemical composition (chemistry) is critical for advancing materials science, yet traditional spectroscopic techniques can be slow and damaging to sensitive samples. Four-dimensional scanning transmission electron microscopy (4D-STEM) captures detailed diffraction patterns across scanned regions, providing rich structural information, while energy dispersive X-ray spectroscopy (EDS) offers complementary chemical data. In this work, we develop a machine learning framework that predicts EDS spectra directly from 4D-STEM diffraction patterns, reducing beam exposure and acquisition time. A convolutional neural network (CNN) accurately infers elemental compositions, particularly for elements with strong diffraction contrast or higher concentrations, such as Oxygen and Tellurium. Both extrapolation and interpolation strategies demonstrate consistent performance, with improved predictions when additional structural context is available. Visual and cross-correlation analyses confirm the model's ability to capture global and local compositional trends. This approach establishes a data-driven pathway to non-destructive, high-throughput materials characterization.