Employing neural density functionals to generate potential energy surfaces.

Area of Science:

• Quantum Chemistry
• Computational Chemistry
• Machine Learning

Background:

• The integration of machine learning (ML) with quantum chemistry is advancing density functional theory (DFT).
• DeepMind's Deep Learning model (DM21) shows state-of-the-art performance, but its application in quantum computations remains underexplored.
• Existing literature lacks studies applying DM21 for generating potential energy surfaces (PES).

Purpose of the Study:

• To demonstrate the generation of potential energy surfaces (PES) using machine-learned density functional approximation (ML-DFA).
• To present the first application of DeepMind's pretrained DM21 neural networks for generating ML-DFA-PES.
• To analyze the long-range behavior of DM21-based PES and compare them with established DFT functionals and CCSD(T).

Main Methods:

• Utilizing pretrained DM21 neural networks within a TensorFlow framework to infer exchange-correlation potentials from electron densities.
• Computing self-consistent field (SCF) energies at various molecular geometries.
• Generating 2D potential energy surfaces (PES) by plotting SCF energies against relevant coordinates.
• Implementing the ML-DFA-PES method using PySCF and DM21 in open-source Python code.

Main Results:

• Successfully generated ML-DFA-PES for C4H8, H2O, H2, and H2+ using the DM21m model with a cc-pVDZ basis set.
• Analyzed the long-range behavior of the generated PES, assessing their descriptive capabilities.
• Compared the DM21-based PES with those obtained from popular DFT functionals (b3lyp, PW6B95) and CCSD(T).

Conclusions:

• The study establishes a novel ML-DFA-based computational method for generating potential energy surfaces.
• The pretrained DM21 neural network shows promise for accurate PES calculations in quantum chemistry.
• This work contributes an open-source implementation for ML-DFA-PES generation, facilitating future research.