Research

Most of my work rests on one idea: an atom’s valence electron cloud can be represented as a quantum Drude oscillator (QDO), defined by just a frequency \(\omega\), a mass \(\mu\) and a charge \(q\). The win is that response and interaction properties then follow analytically, without an electronic-structure calculation.

Two results make the model practical. In the optimized QDO (OQDO) parametrization, all three parameters are fixed from only the static dipole polarizability and the \(C_6\) coefficient — giving accurate polarization potentials and multipolar dispersion coefficients across the periodic table. Building on that, I derived a universal pairwise vdW potential whose functional form comes from the QDO model itself and which again needs only \(\alpha_1\) and \(C_6\) per element. It has since been adopted as the long-range module in machine-learned force-field frameworks including SO3LR and FeNNol.

I also led a community review synthesizing what QDO models can and cannot do, written with colleagues across the field.

Density-functional approximations are usually judged on energies, and long-range dispersion is added afterwards as a correction that leaves the density untouched. That is reasonable for small molecules, but not necessarily for large, polarizable systems — exactly the question I studied.

Across more than 30 molecular and supramolecular systems, benchmarked against carefully converged CCSD−HF reference density differences, dispersion turns out to polarize electron densities substantially, shifting long-range electrostatic potentials by up to 4 kcal/mol. That is a non-trivial coupling between dispersion and electrostatics — two contributions that are routinely treated as separable. This has consequences for any density-derived quantity, from electrostatic potentials to density-based bonding indicators.

There is also an uncomfortable implication for the functionals themselves. In the systems where polarization matters most, an exchange–correlation approximation evaluates the energy at a slightly wrong density, and how much that costs in supramolecular complexes and proteins is what I would eventually like to find out. The hard part is disentangling the two contributions: semilocal functionals already capture some dispersion at intermediate range, by an amount that is difficult to pin down, and that overlap is normally hidden inside an empirically fitted damping function.

Machine-learned force fields have largely solved short-range accuracy, but they inherit a long-standing weakness at long range, where dispersion and electrostatics dominate. Physically grounded, transferable models are the natural fix — and they only need to be derived once.

The universal vdW potential from my QDO work serves exactly this role: it is the long-range dispersion module inside SO3LR, which couples the SO3krates neural network for semilocal interactions with universal pairwise terms for short-range repulsion, electrostatics and dispersion, enabling stable simulations of large and diverse biomolecular systems. In MBD-ML, machine learning supplies the inputs that the nonlocal many-body dispersion (MBD-NL) formalism normally requires from an electronic-structure code, freeing one of the most accurate available dispersion treatments from that dependency.

Both are collaborative projects in which I contributed the long-range dispersion component and expertise in physics rather than leading the work — but they are the clearest evidence that the models in the sections above survive contact with real simulations.

Before moving to noncovalent interactions, I worked on the electronic and mechanical properties of nanomaterials in Moscow. I showed that ultrasmall diamond nanoparticles can be stiffer than bulk diamond depending on their size and shape, and proposed an unambiguous definition of incompressibility at the nanoscale. Separately, I modeled graphene oxide patterned by pristine graphene nanoroads and computed the resulting electronic structure. Several of these projects provided the theoretical counterpart to experimental studies, including h-BN/Al nanocomposite UV photodetectors.

This is not my current direction, but it is where I learned solid-state physics, density-functional theory in practice, and how to work with experimentalists. And who knows, maybe one day I will revisit these topics — with a broader perspective and fresh ideas.