Material discovery by combining stochastic surface walking global optimization with a neural network.

Area of Science:

• Computational Materials Science
• Materials Discovery
• Theoretical Chemistry

Background:

• Predicting new materials from theory is challenging due to complex potential energy surfaces (PES) and inefficient sampling.
• Accuracy of PES and sampling efficiency are major bottlenecks in computational materials science.

Purpose of the Study:

• To introduce a novel "Global-to-Global" approach for accelerated material discovery.
• To combine global optimization with neural network (NN) techniques for efficient PES exploration.
• To demonstrate the SSW-NN method's capability in identifying new material phases and assessing their stability.

Main Methods:

• Developed the stochastic surface walking (SSW) method for massively parallel global optimization.
• Utilized atom-centered NN to fit a multi-dimensional global PES from generated data.
• Applied SSW exploration on the analytical NN PES for large systems to determine thermodynamic and kinetic stability.
• Focused on simultaneous energy, force, and stress NN training for improved accuracy.

Main Results:

• Successfully generated a global training data set using parallel SSW.
• Developed an improved NN training procedure for accurate multi-dimensional global PES.
• Identified two new TiO2 porous crystal structures with thermodynamic stability comparable to rutile.
• Demonstrated kinetic stability for one new TiO2 phase via SSW pathway sampling.

Conclusions:

• The SSW-NN method provides an efficient and predictive platform for large-scale computational material screening.
• This approach overcomes bottlenecks in PES accuracy and sampling for material discovery.
• The method is a general tool for material simulation, enabling accelerated discovery of functional materials.