A Cluster-Based Deep Learning Model Perceiving Series Correlation for Accurate Prediction of Phonon Spectrum.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Spectral properties are crucial for understanding transport phenomena and excitation responses in materials.
• Existing machine learning models struggle to accurately predict spectral properties due to difficulties in perceiving series correlations.

Purpose of the Study:

• To develop a novel machine learning model capable of accurately predicting the phonon density of states (PDOS) spectrum for crystal materials.
• To address the limitations of current machine learning approaches in capturing intrinsic series correlations within spectral data.

Main Methods:

• Developed a cluster-based series graph networks (CSGN) model grounded in the dynamical theory of crystal lattices.
• Constructed a multiple atomic cluster representation to effectively capture diverse vibrational modes.
• Employed a mixture Gaussian process and dynamic time warping mechanism for projecting from atomic clusters to the PDOS spectrum.

Main Results:

• Achieved accurate predictions for complex spectra, including those with multiple or overlapping peaks.
• Demonstrated the model's high performance attributed to pertinent feature extraction and appropriate similarity evaluation.
• Confirmed the model's ability to naturally perceive structure-property relationships and intrinsic series correlations.

Conclusions:

• The CSGN model offers transferable and interpretable predictions, advancing machine learning applications in spectral property analysis.
• The study highlights the potential of designing machine learning methods inspired by physical mechanisms for materials science.
• CSGN represents a significant step forward in predicting spectral properties of crystal materials.