Michele Ceriotti

Details coming soon

Matthias Rupp

Machine Learning for Quantum Mechanics

Data-driven surrogate models of quantum-mechanical observables have seen increasing success over the last decades. I will highlight some research directions related to Anatole von Lilienfeld's work, including models of potential energy surfaces across chemical space. I will then focus on machine-learning interatomic potentials and their trade-offs, showcasing ultra-fast potentials and their applications as an example.

Joshua Schrier

Details coming soon

Yousung Jung

Details coming soon

Markus Reiher

Details coming soon

Guido von Rudorff

Details coming soon

Mark Tuckerman

Details coming soon

Markus Meuwly

Details coming soon

Tristan Bereau

Chemical compound space at lower resolution

Chemical compound space is usually explored at quantum-mechanical resolution, where every nucleus and electron is resolved. Here we explore the same space at lower resolution. Coarse-grained models built from a small alphabet of bead types map many molecules onto a single parametrization, a degeneracy that drastically reduces the size of chemical space and makes it possible to screen thermodynamic properties across hundreds of thousands of small molecules. We exploit this hierarchy of resolutions to navigate chemical space by multi-level Bayesian optimization, in which coarse models guide the search at finer resolution. Finally, we quantify what is lost on the way down: recasting coarse-graining as measure transport between spaces of different dimensionality yields the mapping entropy, the information discarded by a many-to-one representation.

Jochen Blumberger

Adventures in Non-adiabatic Dynamics

Scale-bridging techniques have revolutionized our understanding of condensed phase system but these methods are typically confined to the electronic ground state. In my talk I will describe our 10-year long adventure in developing a beyond Born-Oppenheimer (i.e. non-adiabatic) dynamics approach enabling the simulation of electronically excited processes in truly nanoscale molecular systems (10-100 nanometers) and over long time scales (10-100 ps) using coarse grained electronic Hamiltonians. The method will be illustrated with applications to charge transport, excitation energy transport, light-induced charge generation and thermoelectricity generation in organic semiconducting materials.

Daniel Sebastiani

Details coming soon

Enrico Tapavicza

Details coming soon

Clémence Corminboeuf

Details coming soon

John Keith

Better Data, Clearer Mechanisms: DFT-Referenced Coupled Cluster and Quantum Bond Orders Across Chemical Space

Machine learning models for chemistry are only as reliable as their training data, and most large datasets rely on density functional theory (DFT), whose errors are system-dependent and difficult to anticipate. For catalysis and reaction mechanisms, we also want models that offer chemical insight in addition to accurate predictions. I will first show that coupled cluster theory using a Kohn–Sham DFT reference (KS-CCSD(T)) can outperform both DFT and conventional Hartree–Fock-based CCSD(T). In our recent Nature Commun. paper, we showed that this improvement comes from the KS density encoded into the Fock matrix rather than from the KS orbitals themselves. KS references yield near-chemical accuracy for main-group and transition-metal thermochemistry, they recover the notoriously challenging Cr₂ potential energy surface, and they enable a low-cost diagnostic for multireference character. Extending this approach with DLPNO-CCSD(T) to bond dissociation energy datasets for inorganic compounds, we find the same trends. I will then turn to interpretability and use quantum-mechanical bond orders that are evaluated across stratified subsets of QM9 and the Transition1x reaction dataset. We show how bond orders can be used to read and interpret reaction mechanisms across chemical space. Finally, I will briefly show progress in a family of bond-order-based semi-empirical models (BEBOP2, ZPE-BOP, and QBOP, with KIE-BOP in development) that connect these descriptors to predictive, physically transparent models.

Denis Andrienko

Tensorial Molecular Property Predictions with Equivariant GNNs

We propose a first-principles-based parameterization procedure for atomic polarizability tensors and scalars, validated on a set of small organic molecules with conjugated building blocks. To overcome the computational cost of ab initio calculations, we train a message-passing graph neural network to predict polarizability parameters, enabling efficient and scalable parameterization. This approach imposes no additional computational cost during simulations and provides a clear diagnostic criterion for identifying cases where polarizable force-field models fail to accurately describe molecular polarizability.

Stefan Chmiela

Details coming soon