Effective optimization of atomic decoration in giant and superstructurally ordered crystals with machine learning

Area of Science:

• Materials Science
• Computational Chemistry
• Crystallography

Background:

• Complex crystal structures often exhibit mixed chemical occupancy on Wyckoff sites, complicating accurate atomic modeling.
• Disordered or mixed occupancy can lead to superstructural ordering, significantly increasing unit cell size and computational cost.
• Predicting atomic arrangements in such systems is crucial for understanding material properties and designing new materials.

Purpose of the Study:

• To develop and validate a computational method for predicting optimal atomic decorations in crystals with complex geometries and mixed chemical occupancy.
• To leverage crystal graph convolutional neural networks (CGCNN) and site permutation search (SPS) for efficient and accurate prediction of atomic arrangements.
• To demonstrate the capability of the method in handling both disordered/mixed occupancy and superstructural ordering.

Main Methods:

• Utilized a crystal graph convolutional neural network (CGCNN) to predict the energetic ordering of different atomic decorations for a given chemical composition.
• Implemented a site permutation search (SPS) optimization algorithm, incorporating Monte Carlo moves, simulated annealing, and basin-hopping techniques.
• Employed the CGCNN's energy landscape to guide the SPS in finding the most stable atomic configurations on fixed crystalline geometries.

Main Results:

• The CGCNN-SPS approach accurately predicted atomic decorations for known compounds like Rb8Ga27Sb16 and Cs2SnI6, which were not part of the training set.
• The method successfully determined favorable atomic decorations in crystals with mixed or disordered occupancy and superstructural ordering.
• The critical temperature of an order-disorder phase transition in CuZn was accurately probed by analyzing site configuration trajectories.

Conclusions:

• The developed CGCNN-SPS strategy offers a powerful and efficient tool for predicting favorable atomic decorations in complex crystal structures.
• This approach significantly accelerates the analysis of site occupation in materials with mixed or disordered occupancy and superstructural ordering.
• The method provides accurate predictions, enabling faster discovery and design of novel crystalline materials.