Paul Zimmerman, professor of Chemistry and researcher at the University of Michigan, has introduced a novel method to enhance the accuracy of quantum chemistry simulations. This approach addresses a longstanding challenge in computational chemistry: developing a universally applicable exchange-correlation (XC) functional for density functional theory (DFT).
Kanungo, B., Hatch, J., Zimmerman, P. M., & Gavini, V. (2025). Learning local and semi-local density functionals from exact exchange-correlation potentials and energies. Science Advances, 11(38). https://doi.org/10.1126/sciadv.ady8962
Density functional theory simplifies the complex many-electron Schrödinger equation by focusing on electron density rather than individual electron interactions. This reduction allows for the simulation of larger systems with reduced computational resources. However, the accuracy of DFT heavily depends on the choice of the XC functional, which approximates how electrons interact within a system. Despite extensive research, a universally accurate XC functional has remained elusive.
Paul Zimmerman, professor of Chemistry at the University of Michigan, stated,
“Many-body theories give us the right answer for the right reason, but at an unreasonable computational cost. Our team has translated many-body results into a simpler, faster form that retains most of its accuracy”.
The University of Michigan team, led by Professors Vikram Gavini and Paul Zimmerman, has developed a machine learning-based method to derive a more accurate XC functional. By training neural networks on data obtained from quantum many-body calculations of small atoms and molecules, they have created a functional that improves upon existing approximations.
Their approach utilizes exact electron densities, XC energies, and XC potentials derived from configuration interaction calculations. By solving the inverse Kohn–Sham problem, they determine the corresponding XC potentials and energies, which are then used to train the neural network. The resulting functionals, termed NNLDA and NNGGA, demonstrate significant improvements in total energies and densities compared to traditional functionals. Notably, the NN-based GGA functional achieves accuracy comparable to the higher-rung SCAN meta-GGA functional.
This advancement holds significant promise for various applications in chemistry and materials science. The improved XC functional can lead to more accurate simulations of molecular structures, reaction pathways, and material properties. Such enhancements are crucial for the development of new materials, drug discovery, and the design of efficient catalysts.
The team’s methodology also offers a pathway for systematically improving XC functionals. By expanding the training dataset to include a broader range of atoms, molecules, and solid-state systems, future functionals can achieve even higher accuracy. Additionally, integrating information about electron orbitals could further refine the simulations, though this would require more extensive computational resources.
The integration of machine learning into quantum chemistry represents a significant step forward in computational methods. The University of Michigan’s development of a more accurate XC functional exemplifies how data-driven approaches can address complex challenges in science and engineering. As computational power continues to grow, such innovations will likely play a pivotal role in advancing our understanding and manipulation of molecular systems.
Adrian graduated with a Masters Degree (1st Class Honours) in Chemical Engineering from Chester University along with Harris. His master’s research aimed to develop a standardadised clean water oxygenation transfer procedure to test bubble diffusers that are currently used in the wastewater industry commercial market. He has also undergone placments in both US and China primarely focused within the R&D department and is an associate member of the Institute of Chemical Engineers (IChemE).
