Predictive Models in Chemical Bonding

The element arsenic is known for its toxicity in its elemental form and in both oxidation states As(III) and As(V). Arsenic in natural water systems exists in both organic [e.g., arsanilic acid NH2PhAs(O)(OH)2] and inorganic [arsenite (AsO33-, As(III)), arsenate (AsO43-, As(V)] forms, whereby the inorganic anions are the more common and dangerous forms which are the reason for increasing water contamination today. In our study, we synthesized, crystallized, and measured high-resolution X-ray diffraction data sets of several arsenic oxides and arsenic ylides. We used complementary bonding analysis based on quantum crystallography and based on computational chemistry to understand the nature of chemical bonding within such compounds, with a focus on the role of the As atom, e.g. in terms of its increased atomic polarizability.
In addition, we used the same chemical bonding descriptors to explore intermolecular bonding between As(III) as well as As(V) and their respective environments. As(III) species can act ambivalently as Lewis acids or bases since they possess a lone pair unlike As(V) species which can only act as Lewis acids. However, As filters have so far only been constructed around the assumption that As acts as Lewis acid, although As(III) compounds are the more toxic ones. Hence, here we compare the dualistic intermolecular interaction modes of As(III) compounds to those in As(V) species to facilitate work on As(III) filters for freshwater.
The high toxicity of As compounds is an indicator for their strong interactions with biological environments. Therefore, arsenic has long been used therapeutically as well. Here, we focus on the most controversial historic arsenic-based drug atoxyl (Fig. 1) to explore its characteristics using complementary bonding analysis as well. Finally, we compare it with the previously studied toxic systems to better understand the role of the As atom, even if it is not involved in any intermolecular interactions.

Singlet fission requires a delicate energetic alignment between the lowest singlet and triplet excited states, typically expressed as E(S₁)/E(T₁) > 2. Achieving this condition in compact organic chromophores remains challenging because small π-conjugated systems often exhibit strongly coupled singlet and triplet energetics.
Here, we present a resonance-structure-based strategy for tuning excited-state energies in fused heteroaromatic systems. The approach relies on heteroatom placement as a chemical handle to select specific resonance structures according to the Glidewell–Lloyd rule, which favours the localisation of 4n + 2 π-electron circuits in the smallest available ring. Previous studies on indole and benzoborepine isomers have shown that heteroatom connectivity controls local aromaticity patterns and relative ground-state stability.

We extend this concept to the design of singlet-fission chromophores using isoindole-type frameworks. In these topologies, heteroatom-induced resonance selection reduces ground-state aromatic stabilisation in the six-membered ring, while triplet excitation restores a degree of local Hückel aromaticity in this ring despite the globally Baird-antiaromatic 10π-electron framework. This topology-driven redistribution of aromaticity localises the triplet spin density mainly on the five-membered ring and selectively stabilises T₁.

In contrast, S₁, governed by the same frontier-orbital topology, remains sufficiently high to approach the 2E(T₁) energetic condition required for singlet fission. As a result, isoindole and chalcogen-containing isoindole analogues display enhanced E(S₁)/E(T₁) ratios, approaching or exceeding the singlet-fission threshold.

These results establish heteroatom-guided resonance selection, interpreted through the Glidewell–Lloyd rule, as a transparent design principle for compact singlet-fission chromophores.

This project deals with water molecules in condensed matter. OH vibrational frequencies, 1H NMR chemical shifts and XPS-measured O1s binding energies are widely used experimental techniques to elucidate structure and bonding, and for all three techniques, the spectroscopic signals relate to electron distributions, atomic charges and electronic shielding. At the same time, the physical phenomena behind these spectroscopies are very distinct. So do they respond in a similar fashion to surrounding perturbations, or not? Elucidating this question is Goal No. 1.

I will address that question from a particular perspective, namely by comparing machine-learned ‘structure=>spectra’ models using a number of different structural (and some physical) descriptors, ranging from very simple ones (one H-bond distance) to very complex ones (like the elaborate MACE1 foundation model). Can the simplest structural fingerprints capture the effect of the surrounding molecules on the spectra? And at the other end of the complexity scale – can the models there provide any insight? Elucidating these questions is Goal No. 2.

For the most advanced model in this study (MACE), we also compare some of the electronic features that are now incorporated in the newest MACE versions and exemplify how they compare with results from quantum-mechanical electronic structure calculations. That’s Goal No. 3.

The efficient exploration and screening of chemical reaction space require machine-learning (ML) models capable of making accurate predictions both within and across reaction families. However, when only small, high-quality, task-specific datasets are available, conventional ML architectures often suffer from limited accuracy and poor generalizability. One strategy for addressing this challenge is to augment the models with physically motivated quantum-mechanical (QM) descriptors that encode bonding and electronic-structure information. This approach has been shown to enable robust and data-efficient predictions of chemical reactivity, even when only a few hundred training data points are available.
Unfortunately, generating QM descriptors is computationally demanding, which limits the practicality of this strategy for truly large-scale or high-throughput screening. Surrogate models that predict such descriptors directly from molecular graphs provide a promising alternative. When combined with large, precomputed QM datasets from the literature, these surrogate models can support generalizable screening pipelines that require minimal task-specific training data and incur only marginal computational cost.
In this talk, I will present our recent work in this area, focusing on results showing that downstream predictive models often perform better when they use the internal hidden representations of the surrogate model rather than the predicted descriptors themselves. An exception arises when the descriptor set is closely aligned with the mechanistic physics of the target reaction class. These latent embeddings encode rich and transferable chemical information, making them a powerful foundation for data-efficient predictive chemistry.
Finally, I will discuss how these pre-trained hidden representations can also be used to substantially accelerate Bayesian optimization campaigns.

The development of molecular materials with tailored electronic structures is central for advancing next-generation technologies in catalysis, energy conversion and storage, as well as quantum information. Among emerging electronic materials, electrides stand out as a fundamentally unconventional class of compounds characterized by the presence of excess electrons localized in regions of space not associated with any atomic nuclei, challenging classical concepts of chemical bonding. Their unique electron localization endows them with exceptional properties, including strong reducing power and pronounced nonlinear optical responses, making them highly attractive for functional materials design.

To date, all experimentally confirmed electrides are solid-sate systems, where excess electrons are stabilized within well-defined cavities. In contrast, molecular electrides remain largely unexplored and experimentally elusive due to the intrinsic difficulty of stabilizing non-nuclear localized electrons within discrete molecular frameworks.
In this talk, I will present transition metal (TM) building blocks identified as key candidates for the design of molecular electrides. Guided by a previously established computational strategy, we find that triatomic TM units inherently exhibit electride character, with non-nuclear attractors (NNA) localized at the center of the metal framework. These findings position TM triads as minimal and robust motifs for realizing molecular electrides.