The loss landscape of powder X-ray diffraction-based structure optimization is too rough for gradient descent

Digital Discovery 2026 Open access

Read full text
Authors: Nofit Segal, Akshay Subramanian, Mingda Li, Benjamin Kurt Miller, Rafael Gómez‐Bombarelli
DOI: 10.1039/d6dd00017g
Cited by: 0

Abstract

While potential energy surfaces offer smooth convergence, we show that XRD similarity metrics create a highly non-convex, ill-posed loss landscape. This rugged topology severely complicates gradient-based crystal structure optimization.

Fields of Study & Subjects

X-ray Diffraction in Crystallography Machine Learning in Materials Science Advanced NMR Techniques and Applications

Publication details

Journal / source
Digital Discovery
Volume
5
Issue
4
Pages
1590–1599
Publication date
1 January 2026
Language
EN