DeePKS

DeePKS is a machine-learning (ML) aided density funcitonal model that fits the energy difference between highly accurate but computationally demanding method and effcient but less accurate method via neural-network. Common high-precision methods include hybrid functionals or CCSD-T, while common low-precision methods are LDA/GGA.

As such, the trained DeePKS model can provide highly accurate energetics (and forces/band gap/density) with relatively low computational cost, and can therefore act as a bridge to connect expensive quantum mechanic data and machine-learning-based potentials. While the original framework of DeePKS is for molecular systems, please refer to this J. Phys. Chem. A 126.49 (2022): 9154-9164 for the application of DeePKS in periodic systems.

Detailed instructions on installing and running DeePKS can be found on this website. The DeePKS-related keywords in INPUT file can be found here. An example for training DeePKS model with ABACUS is also provided. For practical applications, users can refer to a series of Notebooks. These Notebooks provide detailed instructions on how to train and use the DeePKS model using perovskite as an example. Currently, these tutorials are available in Chinese, but we plan to release corresponding English versions in the near future.

Note: DeePKS calculations can only be performed by the LCAO basis.