Machine-learned density functional based quantum chemical computations for ethane: performance of DeepMind 21 on potential energy surface and molecular properties

Abstract

Context

Machine learning (ML) has proven to be a promising method in quantum chemistry calculations. Researchers at Google DeepMind established the superior performance of a machine-learned functional DeepMind 21 (DM21), the neural network-based functionals which emerged as a powerful tool for developing exchange–correlation energy approximation in density functional theory (DFT), to accurately explain the charge delocalization and strong correlation. While many researchers have cited the work of DeepMind, still there remains a paucity of research publications that perform algorithmic computations with DeepMind’s AI model for quantum sciences. To address this lacuna, this paper investigates quantum chemistry algorithmic computations of potential energy surface and analysis properties of ethane molecule (C₂H₆) by employing the machine learning model to predict exchange–correlation potential. This paper utilizes neural density functional, DeepMind 21 to compute the dipole moment, molecular orbitals (HOMO/LUMO), and long-range interactions. Our computations were based on DM21m TensorFlow neural network model-based prediction of exchange–correlation potential, and then using this prediction to compute self-consistent field energies using PySCF Python package with cc-pVDZ Dunning dual basis set. The accuracy of the DM21 functional was rigorously assessed through comparison with conventional DFT methods (B3LYP, PW6B95) and the reference CCSD(T) standard. The findings indicate that the DM21 functional yields result in close agreement with the CCSD(T) benchmark energies and established literature values, confirming its efficacy for PES generation and quantum chemical computation in systems like ethane. This investigation demonstrates the suitability of the deep learning density functional for quantum science computations of ethane molecule for the first time.

Methods

In this method, the potential energy surfaces and properties (dipole moment, molecular orbitals) are computed using machine-learned density function approximation using pretrained deep learning models provided by DeepMind 21 researchers. By inserting deep learning inference in density functional theory (DFT) with a pretrained neural network, self-consistent field (SCF) energy at different geometries along the coordinate of interest is computed, and hence potential energy surfaces are computed and obtained. For a given specified molecular geometry, the algorithm computes the electron density vector, which is then used as a machine learning feature input for a pretrained DM21 deep learning model to predict the exchange–correlation. This methodology was implemented in a Python source code using frameworks such as PySCF and DM21.