Accelerating CALYPSO structure prediction by data-driven learning of a potential energy surface.

Area of Science:

• Computational materials science
• Chemical physics
• Machine learning applications

Background:

• Ab initio structure prediction methods are crucial for materials discovery but computationally expensive for large systems.
• Density functional theory (DFT) calculations limit the scale of current structure prediction methods.

Purpose of the Study:

• To develop accelerated schemes for predicting the structures of large atomic clusters.
• To combine machine learning (ML) potentials with the CALYPSO structure prediction method.

Main Methods:

• Developed two ML-potential-based acceleration schemes for structure prediction.
• Pre-constructed ML potentials or trained them on-the-fly to replace DFT calculations.
• Applied the methods to medium- and large-sized boron clusters.

Main Results:

• Successfully reproduced experimental structures for B36 and B40 boron clusters.
• Proposed a putative global minimum structure for the B84 boron cluster.
• Achieved a computational cost reduction of 1-2 orders of magnitude compared to DFT-based methods.

Conclusions:

• The combination of ML techniques and structure prediction methods offers a viable route for large system analysis.
• Accelerated methods enable efficient structure prediction and materials discovery for larger atomic systems.