Machine Learning Diffusion Monte Carlo Energies.

Area of Science:

• Computational materials science
• Machine learning in chemistry
• Quantum mechanical modeling

Background:

• Predicting energies from diffusion Monte Carlo (DMC) calculations is computationally expensive.
• Accurate energy predictions are crucial for understanding material properties and chemical reactions.
• Existing methods often require large datasets, limiting their applicability.

Purpose of the Study:

• To develop and compare machine learning methodologies for predicting DMC energies with limited data.
• To assess the performance of kernel ridge regression (KRR) and voxel deep neural networks (VDNNs).
• To evaluate the generalizability of KRR models across different material systems and properties.

Main Methods:

• Voxel deep neural networks (VDNNs) were used to predict DMC energy densities from DFT electron densities.
• Kernel ridge regression (KRR) was employed to predict atomic contributions to DMC energies using atomic environment vectors.
• Methodologies were tested on graphene, Stone-Wales defects, and liquid water systems.

Main Results:

• KRR demonstrated superior performance compared to VDNNs, gradient boosted decision trees, random forest, Gaussian process regression, and multilayer perceptrons.
• KRR models achieved higher accuracy than Kohn-Sham DFT, with mean absolute errors below chemical accuracy.
• KRR models showed good generalizability for predicting energy barriers and total energies in diverse systems.

Conclusions:

• KRR is a highly effective method for predicting DMC energies with small datasets.
• The developed KRR models offer a computationally efficient and accurate alternative to traditional DFT and DMC calculations.
• This approach has significant potential for accelerating materials discovery and design.