Leverage electron properties to predict phonon properties via transfer learning for semiconductors.

Area of Science:

• Materials Science
• Computational Materials Science
• Solid State Physics

Background:

• Predicting phonon properties is crucial for materials design in thermoelectrics and electronics.
• Phonon properties are generally more challenging to obtain than electron properties.
• Leveraging electron properties can accelerate the discovery of materials with desired thermal and electronic characteristics.

Purpose of the Study:

• To demonstrate the efficacy of transfer learning (TL) in predicting phonon properties.
• To improve machine learning models for predicting phonon bandgap, group velocity, and heat capacity.
• To explore the use of readily available electron properties to infer complex phonon behavior.

Main Methods:

• Trained machine learning models on electronic bandgaps of 1245 semiconductors.
• Applied transfer learning (TL) to retrain models with a smaller dataset (124 data points) for phonon property prediction.
• Validated the performance of TL models against directly trained models and external datasets.

Main Results:

• TL significantly reduced prediction errors for phonon bandgap (65%), group velocity (14%), and heat capacity (54%) compared to direct training.
• TL models demonstrated strong predictive power even when trained on limited data.
• The study confirmed TL's ability to utilize related, less accurate, proxy properties for improved target property prediction.

Conclusions:

• Transfer learning offers a powerful and efficient method for predicting diverse phonon properties.
• This approach accelerates materials discovery by leveraging existing electronic property data.
• TL's capability to use proxy properties broadens its applicability in materials informatics and data-driven design.