Approaching exact density-functional theory: deep-learnt exchange-correlation potential

Professor Chen, Guanhua   (Principal Investigator (PI))

36

2021-01-01

599861

Approaching exact density-functional theory: deep-learnt exchange-correlation potential

convolutional neural network, deep learning, Density-functional theory, electron density, exchange-correlation potential

Chemical Sciences,Others - Physical Sciences

Physical Sciences (P)

17309620

General Research Fund (GRF)

2020

On-going

1 Building upon our earlier success on small molecules and ions made of H and He, we will extend 3D-CNN to include methane. CCSD(T)/aug-cc-pVQZ will be used to calculate the electron density function of methane. Methane will be distorted to form a series of structures, each of which will be calculated to obtain the high precision electron density. The resulting electron densities will be used to calculate their corresponding effective exchange-correlation potential via Wu-Yang method. The 3D-CNN will then be trained and validated for methane. Finally, KS-DFT/NN calculation will be carried out on methane to confirm our KS-DFT/NN applicable to methane. 2 We will also extend 3D-CNN to treat water molecule. CCSD(T)/aug-cc-pVQZ will be used to calculate the electron density function. Water molecule will be distorted to form a series of structures, each of which will be calculated to obtain the high precision electron density. The 3D-CNN will then be trained and validated for water molecule. Finally, KS-DFT/NN calculation will be carried out on methane to confirm our KS-DFT/NN applicable to water molecule. Zero-force condition will be enforced. 3 To extend the 3D-CNN and KS-DFT/NN to organic molecules. Building upon the success of H2, He2, HeH+, CH4 and H2O, we are ready to extend our method to more complex organic molecules containing H, C and O. The dataset will be generalized to include these complex molecules or ions. As only quasi-local electron density is required, we expect that the sizes of the molecules and ions required for the dataset are limited, for instance, containing less than 10 atoms. The 3D-CNN that maps quasi-local electron density and its local exchange-correlation potential will be duly trained, validated and developed for most molecules and ions that are encountered in research and industries. 4 We will extend further our 3D-CNN and KS-DFT/NN to treat the general organic molecules containing H, C, O, N, S, F and Cl. CCSD(T)/aug-cc-pVQZ will be used to calculate the electron density, and supercomputers will be employed. Large amount of GPUs will also be used to speed up the training and validation of 3D-CNN as dataset is expected to be huge. Sufficient computing resources will be deployed.