Electronic structure reading group - Matija Medvivović
Event details
| Date | 28.10.2026 |
| Hour | 16:00 › 17:00 |
| Speaker | Matija Medvivović |
| Location | |
| Category | Conferences - Seminars |
| Event Language | English |
Optimizing and inverting Kohn–Sham potentials with "multipole splats"
Local effective potentials are central to Kohn–Sham density functional theory, but computing them remains numerically challenging. Both the optimized effective potential problem for orbital-dependent functionals and the inverse problem of reconstructing a potential from a prescribed density yield ill-conditioned problems in finite orbital bases. Accurate energies and densities can coexist with unphysical potentials, making the physics-informed representation and regularization of the potential essential parts of the problem.
In this talk, I will present a common framework for these two problems as nonlinear optimization over parametrized local Hamiltonians. The optimized effective potential follows from variational energy minimization, while density inversion admits a statistical learning interpretation through quantum relative entropy, recovering the Wu–Yang objective in the zero-temperature limit. We propose "multipole splats”, a class of expressive and nonlinearly parameterized potentials. This construction enforces the required asymptotic decay through a charge constraint and retains an analytical effective Hamiltonian matrix in standard Gaussian orbital bases. Direct first-order optimization, supplemented by a physically motivated regularizer, avoids explicit inversion of the Kohn–Sham response. I will demonstrate the approach for exact exchange, hybrid functionals, and inversion of correlated reference densities. Applications include spatially resolved exchange-correlation potential errors during molecular dissociation and the reconstruction of atomic Rydberg series. These examples illustrate how local potentials expose spatially-resolved differences between approximations.
Local effective potentials are central to Kohn–Sham density functional theory, but computing them remains numerically challenging. Both the optimized effective potential problem for orbital-dependent functionals and the inverse problem of reconstructing a potential from a prescribed density yield ill-conditioned problems in finite orbital bases. Accurate energies and densities can coexist with unphysical potentials, making the physics-informed representation and regularization of the potential essential parts of the problem.
In this talk, I will present a common framework for these two problems as nonlinear optimization over parametrized local Hamiltonians. The optimized effective potential follows from variational energy minimization, while density inversion admits a statistical learning interpretation through quantum relative entropy, recovering the Wu–Yang objective in the zero-temperature limit. We propose "multipole splats”, a class of expressive and nonlinearly parameterized potentials. This construction enforces the required asymptotic decay through a charge constraint and retains an analytical effective Hamiltonian matrix in standard Gaussian orbital bases. Direct first-order optimization, supplemented by a physically motivated regularizer, avoids explicit inversion of the Kohn–Sham response. I will demonstrate the approach for exact exchange, hybrid functionals, and inversion of correlated reference densities. Applications include spatially resolved exchange-correlation potential errors during molecular dissociation and the reconstruction of atomic Rydberg series. These examples illustrate how local potentials expose spatially-resolved differences between approximations.
Practical information
- Informed public
- Free
Organizer
- Noé Blassel
Contact
- Noé Blassel