Node transfer for multi-fidelity and multimodal machine learning for predicting experimental bandgaps.

Area of Science:

• Materials Science
• Computational Materials Science
• Machine Learning

Background:

• Bandgap is a critical material property influencing electronic and optical behavior.
• Machine learning (ML) shows promise for predicting experimental bandgaps, but accuracy improvements are needed.
• Integrating diverse data sources is key to enhancing ML prediction capabilities.

Purpose of the Study:

• To develop an advanced ML framework for accurate bandgap prediction.
• To introduce and evaluate a novel information-fusion strategy called node transfer.
• To enable efficient materials design by predicting bandgaps using only chemical composition.

Main Methods:

• Constructed a multi-fidelity and multimodal ML framework.
• Integrated heterogeneous data from first-principle calculations and X-ray diffraction (XRD) spectra.
• Proposed and implemented the node transfer information-fusion strategy, comparing it with the Δ-learning strategy.

Main Results:

• Node transfer consistently outperformed Δ-learning across various benchmarks.
• The best model, utilizing XRD-based and pre-trained encoded descriptors, achieved a mean absolute error of 0.258 eV.
• This represents a 26.3% reduction in error compared to the single-fidelity baseline.

Conclusions:

• The proposed node transfer strategy effectively integrates heterogeneous data for improved ML predictions.
• The developed ML models accurately predict bandgaps using only chemical composition, facilitating rapid materials discovery.
• This approach significantly advances the predictive power of ML in materials science.