Crystal Structure Representation for Neural Networks using Topological Approach

Area of Science:

• Materials Science
• Computational Chemistry
• Crystallography

Background:

• Predicting crystal properties is crucial for materials design.
• Traditional methods can be computationally intensive.
• Developing efficient predictive models is an ongoing challenge.

Purpose of the Study:

• To develop a new method for predicting physicochemical properties of crystals.
• To utilize crystal topology and artificial neural networks (ANNs) for property prediction.
• To demonstrate the efficacy of this approach across various crystal types.

Main Methods:

• Crystal structures of 268 compounds were analyzed using ToposPro software to obtain their topologies.
• Quotient graphs were employed to identify topological centers and their adjacent atoms.
• Artificial neural networks (ANNs) were trained using the Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm to predict properties.
• The model was trained on properties including molar heat capacity, standard molar entropy, and lattice energy.

Main Results:

• The ANN model achieved a mean absolute percentage error (MAPE) of less than or equal to 8% for the predicted properties.
• The topological approach proved effective for a diverse set of crystal structures, including metals, inorganic salts, and oxides.
• Accurate prediction of key physicochemical properties was demonstrated.

Conclusions:

• Crystal topology serves as a valuable descriptor for predicting physicochemical properties.
• Artificial neural networks, combined with topological analysis, offer a powerful and accurate method for materials property prediction.
• This approach provides a computationally efficient alternative for materials research and development.