Neural Networks Accelerate the Ab Initio Prediction of Solid-Solid Phase Transitions at High Pressures

Area of Science:

• Computational chemistry
• Materials science
• Solid-state physics

Background:

• High-accuracy ab initio calculations like Møller-Plesset perturbation (MP2) are crucial for understanding material properties.
• However, the high computational cost of these methods limits their application to large systems, such as macromolecules.

Purpose of the Study:

• To develop an efficient computational approach for predicting Gibbs free energy, structural characteristics, and phase transitions in solid crystal structures.
• To overcome the limitations of traditional high-accuracy computational methods for large-scale systems.

Main Methods:

• A novel approach combining neural networks and the fragment method was developed.
• This method predicts thermodynamic and structural properties of solid crystals.

Main Results:

• The proposed method achieves prediction accuracy comparable to MP2 calculations.
• It is hundreds of times faster than MP2, significantly reducing computational cost.
• Predicted structures and phase transitions for ice phases IX and XV under extreme conditions showed excellent agreement with MP2 calculations and experimental data.

Conclusions:

• The developed approach accurately and efficiently predicts high-pressure structures and phase diagrams of solid systems.
• It effectively addresses the challenge of high computational cost in high-precision theoretical studies of high-pressure molecular crystals.
• This method has potential applications in materials science and condensed matter physics.