Ab initio structure solutions from nanocrystalline powder diffraction data via diffusion models

Area of Science:

• Materials Science
• Crystallography
• Machine Learning

Background:

• Determining the structure of nanomaterials is a significant challenge in materials science.
• Existing methods struggle with nanoscale object structural determination.

Purpose of the Study:

• To develop a machine learning approach for solving the structure of nanometre-sized objects.
• To utilize generative diffusion models trained on known structures for accurate structural determination.

Main Methods:

• A generative machine learning model, PXRDnet, based on diffusion processes was developed.
• The model was trained on 45,229 known material structures.
• PXRDnet factors measured diffraction patterns and statistical priors on unit cell structures.

Main Results:

• PXRDnet successfully solved simulated nanocrystals as small as 10 Å across 200 materials.
• The model achieved high accuracy, determining structural candidates four out of five times with an average R-factor error of 7%.
• PXRDnet demonstrated capability in solving structures from noisy, real-world experimental diffraction patterns.

Conclusions:

• Data-driven approaches, augmented by theoretical simulations, offer a promising path for solving previously undetermined nanomaterial structures.
• PXRDnet represents a significant advancement in the structural analysis of nanomaterials.
• This method has the potential to accelerate the discovery of new materials.