Chemical Bonding in the Era of Artificial Intelligence

Energy profiles alone miss key aspects of how reactions actually proceed: which bonds break first, how charge flows between atoms, and where the energetic cost of a barrier really comes from. Real-space bonding tools such as QTAIM and IQA can answer these questions, but have so far been applied only to small, hand-picked systems.
We present MARS-ROVER (Reaction Optimization and Virtual Exploration of Reactivity), a module within the MARS platform that automates reaction mechanism discovery using machine-learned force fields (MLFFs). Starting from just reactant and product structures, it generates reaction paths through nudged elastic band calculations, refines saddle points, and validates intrinsic reaction coordinates (IRCs). New intermediates found along the way are fed back automatically, building up complete reaction networks without manual intervention. The toolkit works with modern MLFFs such as SO3LR, MACE, and ANI, and runs on standard laptop hardware. We have applied it to competing E2/SN2 pathways, pericyclic reactions (Diels–Alder, Cope, 1,3-dipolar cycloadditions), organocatalytic cycles, and regio- and stereoselective processes.
We then go beyond energy profiles by analyzing each pathway with real-space descriptors: IQA decomposition to see how electrostatic, exchange, and intra-atomic energies evolve as bonds form and break; delocalization indices and bond orders to track bond reorganization; and atomic charges to follow charge transfer along the reaction coordinate. This combination reveals features invisible to energy-only analysis, such as the degree of synchronicity in multi-bond processes, which atom–atom interactions control barrier heights, and how charge-transfer patterns differ between competing pathways. By pairing fast MLFF exploration with the chemical interpretability of real-space bonding analysis, we can now characterize entire reaction networks at a level of detail that was previously limited to individual case studies.

We pioneered the use of Gaussian Process Regression in atomic potentials. Contrary to other research groups we started with the machine learning (ML) of electrostatics. We covered all 20 amino acids because FFLUX targets protein simulation. Electrostatics manifests itself at long-range and, in order to reach the exact limit, one must work with multipole moments. Next followed accurate ML predictions of the atomic energies using quantum topological atoms, such that a single partitioning method provides all atomic properties. This uniformity guarantee accuracy for the right reasons and hence future-proof work. Sustained in-house software development delivered the training program FEREBUS supported by the pipeline ICHOR and the MD program DL_FFLUX. ML methods such as grey-wolf optimisation are used for the first time with superior performance. The parallellisation of DL_FFLUX enables the simulation of condensed matter, whether molecular crystals or liquid water, all with polarisable electrostatics. We also reached a proof-of-concept replacing Lennard-Jones potentials for both repulsion and dispersion. FFLUX simulations show unprecedented robustness with competitive CPU timings.

When a feed of molecules is sent over a nanoporous solid, they undergo a fascinating journey characterized by various events like adsorption, diffusion, reaction (Figure 1). Eventually after having spent some time in the material, they escape from the lattice either in transformed or changed form. All the sketched events, are characterized by vastly different length and time scales varying from the picosecond to the second/hours and the nano- to the micrometre. Length and time phenomena are deeply intertwined, any change made on the nanometre scale impacts the further time behaviour of the molecules. Furthermore the behavior of an active catalyst is strongly affected by the operating conditions, catalysts also dynamically evolve in time. Modeling such catalysts in action from the molecular level is extremely challenging.

In today’s modeling practice, each of the sketched events is modeled in an isolated way with inconsistent methods, i.e. reactions are modeled with quantum mechanical methods whereas transport phenomena are modeled with classical force fields. In this talk, evidence is given that reaction, diffusion and adsorption are dynamically entangled under realistic time-dependent operating conditions and that we need quantum-accurate dynamics across scales to simulate the true behavior of a catalyst. In this context, machine‑learning potentials (MLPs)—numerical potentials trained on quantum‑mechanical data—are instrumental for extending simulations to much longer length and time scales. Key reactions and intermediates occurring during the CO2 to hydrocarbons and methanol conversion over zeolites are discussed. The results show how MLPs provide quantum‑accurate dynamic insights into individual catalytic steps. However to obtain an overall time perspective of a molecular trajectory, connection with other multiscale kinetic and transport models will be necessary, ultimately enabling a direct connection between detailed molecular events and the full time evolution of industrial catalysts under realistic conditions.