Neural-network-backed evolutionary search for SrTiO₃(110) surface reconstructions.

Area of Science:

• Materials Science
• Computational Chemistry
• Surface Science

Background:

• Determining atomic structures in surface reconstructions traditionally relies on intuition and domain knowledge.
• Evolutionary algorithms are effective for structure searches, but density functional theory (DFT) calculations are computationally expensive for thorough energy landscape exploration.

Purpose of the Study:

• To develop a computationally efficient method for exploring the phase diagram of TiO2 overlayer structures on SrTiO3(110).
• To combine evolutionary algorithms with a surrogate energy model for accelerated structure prediction.

Main Methods:

• Employed the covariance matrix adaptation evolution strategy (CMA-ES) coupled with a neural-network force field (NNFF) as a surrogate energy model.
• Trained the NNFF on SrTiO3(110) 4x1 overlayer structures and tested its transferability on 3x1, 4x1, and 5x1 overlayers.
• Utilized the speedup from the surrogate model to perform exhaustive explorations of the potential energy landscape.

Main Results:

• Successfully demonstrated the transferability of the NNFF across different overlayer structures.
• Identified both previously known and novel low-energy reconstructions for TiO2 on SrTiO3(110).
• Achieved significant computational speedup compared to traditional DFT-based methods.

Conclusions:

• The combination of CMA-ES and NNFF provides an efficient and effective approach for exploring complex surface reconstruction phase diagrams.
• This method accelerates the discovery of stable atomic structures, advancing surface science and materials design.