Coarse-Grained Density Functional Theory Predictions via Deep Kernel Learning.

Area of Science:

• Computational Materials Science
• Quantum Chemistry
• Machine Learning

Background:

• Scalable electronic predictions are crucial for designing soft materials.
• Existing Electronic Coarse-Graining (ECG) methods use deep neural networks (DNNs) but face overfitting and uncertainty issues.
• Deep Kernel Learning (DKL) offers a probabilistic approach to machine learning.

Purpose of the Study:

• To develop a GPU-accelerated Deep Kernel Learning (DKL) framework for Electronic Coarse-Graining (ECG).
• To enable coarse-grained (CG) quantum chemical (QC) predictions using range-separated hybrid density functional theory (DFT).
• To improve the accuracy and efficiency of electronic property predictions for soft materials.

Main Methods:

• Integrated ECG with a GPU-accelerated DKL framework.
• Treated predicted electronic properties as random Gaussian Processes within DKL.
• Incorporated CG mapping degeneracy by learning the distribution of electronic energies.
• Utilized active learning guided by DKL uncertainties for efficient training.

Main Results:

• Achieved a 10⁷ speedup compared to naive all-atom QC calculations.
• Accurately reproduced molecular orbital energies from range-separated DFT.
• Demonstrated efficient training through active learning, sampling diverse configurational spaces.
• Observed comparable performance across different active learning query methods due to feature space overlap.

Conclusions:

• DKL-ECG provides a scalable and accurate approach for CG QC predictions.
• The framework significantly accelerates simulations for soft materials design.
• Active learning enhances training efficiency, although performance is robust across methods for the studied system.