Transformer-based deep learning structure-conductance relationships in gold and silver nanowires

Area of Science:

• Materials Science
• Computational Chemistry
• Nanotechnology

Background:

• Nano-junctions exhibit variable conductance due to stochastic structures, hindering structure-conductance relationship establishment.
• Observing nano-junction structural evolution during conductance measurement is experimentally challenging.

Purpose of the Study:

• To develop a deep learning approach for predicting nano-junction conductance based on atomic structure.
• To establish a reliable method for understanding the structure-conductance relationship in nano-junctions.
• To explore the application of transformer-based neural networks in molecule-scale electronics.

Main Methods:

• Classical molecular dynamics (MD) simulations with neural-network potentials were used to model Au and Ag nanowire stretching.
• A transformer-based neural network was trained to predict the zero-bias conductance from simulated nano-junction structures.
• The model's accuracy, stability, and scalability were evaluated, including its transferability to different materials.

Main Results:

• The transformer network achieved high accuracy in predicting conductance, comparable to ab initio methods but with lower computational cost.
• The model demonstrated excellent stability and scalability, accurately predicting conductance for larger and structurally diverse gold nanowires.
• Simulated conductance histograms closely matched experimental data, and temperature-dependent structural changes were linked to conductance peak shifts.

Conclusions:

• Deep learning, specifically transformer networks, offers an efficient and accurate approach to predict nano-junction conductance and elucidate conduction mechanisms.
• The developed method shows promise for advancing molecule-scale electronics by enabling precise structure-conductance relationship analysis.
• The neural network's transferability facilitates accurate conductance prediction for new materials like silver nanowires with minimal data.