Machine learning climbs the Jacob's Ladder of optoelectronic properties.

Area of Science:

• Computational materials science
• Machine learning applications
• Quantum chemistry

Background:

• Machine learning (ML) models for predicting optoelectronic properties are limited by training datasets calculated using the independent-particle approximation (IPA).
• Higher-accuracy methods like the random phase approximation (RPA) offer better data but are computationally expensive, limiting dataset size.

Purpose of the Study:

• To investigate the effectiveness of transfer learning in improving ML models for optoelectronic property prediction.
• To demonstrate how ML models can be fine-tuned using limited high-fidelity RPA data to overcome limitations of IPA-based datasets.

Main Methods:

• Utilized a graph attention network architecture for ML predictions.
• Trained an initial model on a large dataset of 10,000 independent-particle approximation (IPA) calculations.
• Fine-tuned the model using a smaller dataset of approximately 300 random phase approximation (RPA) calculations.

Main Results:

• Fine-tuning with a small set of RPA data significantly improved prediction accuracy, approaching that of models trained on a large RPA dataset (6000 spectra).
• Transfer learning demonstrated the value of high-fidelity data, even in small quantities, for enhancing optical property predictions.
• The model showed effective generalization to larger unit cells after retraining on RPA data from smaller unit cells, indicating broad scalability.

Conclusions:

• Transfer learning is a viable strategy to enhance ML models for optoelectronic properties by incorporating high-fidelity data.
• Even limited RPA calculations can substantially improve ML model performance and accuracy.
• The developed approach demonstrates scalability and potential for broader applications in materials discovery and design.