Poster Presentation Royal Australian Chemical Institute National Congress 2026

Machine learning the universal density functional through exposure to diverse chemical properties (#220)

Peter Halat 1 , Florencia Stella 2 , Nicole Wang 1 , Russhil Khurana 1 , Ekaterina I Izgorodina 1
  1. School of Chemistry, Monash University, Clayton, Victoria, Australia
  2. Monash Institute of Pharmaceutical Sciences, Parkville, Victoria, Australia

Density Functional theory, developed in the 1960s, dictates that there exists a one-to-one correspondence between the electron density of a chemical system and its ground state energy.1,2 Since then, numerous density functionals have been developed, empowering scientists to calculate energetic properties of their compounds of interest.3

Typical approaches to improve the accuracy of density functionals is to mix in components of wavefunction based methods, such as Hartree-Fock theory or second order Møller-Plesset theory.4 With these wavefunction method components comes an increase in computational cost. As machine learning methods rise in popularity, their applicability for quantum chemistry is being continually tested.

Herein, a neural network is trained to predict the atomization energies and interaction energies from a variety of datasets, using the electron density of inexpensive density functionals as an input. Specifically, the LDA functional was implemented to calculate the densities of molecules belonging to the TAE140 and QM7b datasets of atomization energies. Furthermore, LDA electron densities were provided as inputs to learn from interaction energy databases such as S66, IL174 and SSI. When both interaction and atomization energies were combined into a single dataset, the neural network provided predictions within 1 kcal/mol of reference energies, being highly competitive with previous machine learning approaches and state-of-the-art density functionals.4

The same trained neural network also predicts unseen reaction energies and reaction barriers to high accuracy, and further incorporation of these databases into the training set increases the overall accuracy of the model.

This work acts as a preliminary study into the effectiveness of electron density inputs into neural networks into the prediction of chemical properties, providing a new direction of density functional design, towards a universal computational chemistry model.

  1. Self-consistent equations including exchange and correlation effects. W. Kohn, L. J. Sham, Phys. Rev., 1965, 140, A1133.
  2. Inhomogeneous Electron Gas. P. Hohenberg, W. Kohn, Phys. Rev., 1964, 136, B684.
  3. Fifty years of density-functional theory in chemical physics. A. D. Becke, J.Chem. Phys., 2014, 140, 18A301.
  4. Thirty years of density functional theory in computational chemistry: an overview and extensive assessment of 200 density functionals. N. Mardirossian, M. Head-Gordon, Mol. Phys., 2017, 115, 2315-2372.