Enhancing SchNet-Based Structure Prediction for Doped Clusters via Transfer Learning and Fine-Tuning.

Area of Science:

• Computational chemistry and materials science.
• Application of machine learning in predicting atomic cluster structures.

Background:

• Doped clusters' electronic and magnetic properties are tunable via heteroatoms, crucial for applications.
• Accurate structure prediction is vital for understanding structure-property relationships in doped clusters.
• Existing machine learning (ML) methods face challenges with data demands and model portability for heterogeneous clusters.

Purpose of the Study:

• To develop an efficient ML method for predicting the global minimum structures of doped silicon clusters.
• To address data and computational bottlenecks in predicting heterogeneous cluster structures.
• To establish a standardized ML framework for doped cluster studies.

Main Methods:

• Utilized the SchNet model, a framework suitable for physicochemical tasks.
• Integrated transfer learning by freezing neural network layers and fine-tuning with minimal data.
• Constructed a dataset for EuSi_n (n=3-12) clusters, with energy validation using DFT calculations.
• Optimized the ML model for predicting the structures of EuSi_n clusters.

Main Results:

• The transfer-learned ML model accurately predicted the global minimum structures of EuSi_n (n=3-12) clusters.
• Achieved results consistent with traditional density functional theory calculations.
• Reduced computational time by 54.09% and data requirements by 88.89% compared to the original SchNet model.

Conclusions:

• The proposed method overcomes traditional bottlenecks in doped cluster calculations.
• Demonstrated significant efficiency gains in both computation time and data requirements.
• Provides a new paradigm for machine learning studies on doped clusters.