Bypassing backmapping: Coarse-grained electronic property distributions using heteroscedastic Gaussian processes.

Area of Science:

• Computational Chemistry
• Materials Science
• Statistical Mechanics

Background:

• Coarse-grained (CG) models simplify complex systems but often lose fine-grained electronic details.
• Recovering all-atom properties from CG simulations is crucial for accurate predictions.
• Existing methods like backmapping can be computationally intensive.

Purpose of the Study:

• To develop a flexible and scalable framework for learning electronic property distributions from CG models.
• To quantify the impact of CG model resolution on predictive accuracy.
• To identify CG mapping operators that preserve essential electronic degrees of freedom.

Main Methods:

• Deep kernel learning electronic coarse-graining (DKL-ECG) with approximate Gaussian processes.
• Analysis of heteroscedastic electronic property distributions as a function of CG configuration.
• Examination of CG model resolution effects on predictive distribution certainty.
• Iterative Boltzmann inversion combined with DKL-ECG.

Main Results:

• DKL-ECG successfully learns electronic property distributions from CG models.
• Predictive certainty is limited by CG model resolution, converging to intrinsic physical noise.
• Identified CG mapping operators effectively capture electron-phonon coupling.
• Exact valence electronic density of states (EDOS), including tail behavior, was reconstructed from CG models.

Conclusions:

• DKL-ECG offers a powerful method to learn and recover all-atom electronic properties lost in CG models.
• This approach provides a promising alternative to traditional backmapping techniques.
• The framework allows for a deeper understanding of how CG resolution impacts electronic property prediction.