While artificial intelligence is driving a paradigm shift across many scientific disciplines, few areas showcase its transformative power as clearly as ab initio molecular dynamics (aiMD) simulations. Through so-called machine-learned interatomic potentials (MLIPs) , AI has enabled a dramatic acceleration in the propagation of the equations of motion (typically by 3-4 orders of magnitude), while retaining quantum-level accuracy in describing electronic degrees of freedom.

In condensed matter physics and materials science, MLIPs have emerged as a game-changing approach. Over the past 15–20 years, and especially in the past five, significant advances have been made in the machine learning architectures that underpin these potentials. MLIPs are also poised to revolutionize biomolecular simulations, offering unprecedented accuracy and likely to offer answers to the ongoing challenges posed by empirical classical forcefields.

A rapidly growing application area is in atomistic simulations of chemical reactivity, though this domain presents unique challenges compared to others. There have recently been many proof-of-concepts and ground-breaking applications to reactivity studies in the last few years, and the field has now reached a turning point when further extension would require core developments. Key issues include, but are not limited to, 

• Efficient construction of training datasets: Chemistry poses a "chicken-and-egg" problem: training an MLIP requires representative structures along reactive pathways, but generating such structures is tedious without an already trained MLIP.

• Sampling of chemical space and rare events: Effective exploration of chemical space demands careful identification of collective variables and efficient strategies for sampling rare events.

• Applications beyond ground-state chemistry: Most MLIP-based studies of chemical reactivity have focused on ground-state reactions. However, many important chemical processes involve excited states and light–matter interactions, which are only beginning to be addressed by ML approaches.

• Extensions to biological systems: A frontier area involves coupling MLIPs with classical molecular mechanics (MM) models to simulate chemical reactivity in biological environments. This is analogous to QM/MM methods, but leverages the speed advantages of MLIPs.

• Foundation models vs. task-specific models: A growing trend is the development of general-purpose MLIPs capable of addressing a wide range of chemical systems. This contrasts with the historical focus on system-specific models, which can be more accurate and efficient for targeted applications.

We will have limited available space for participants in addition to the invited speakers, with a priority given to young researchers. We do apologize in advance if we cannot accomodate everyone. Please apply early if possible. We will most likely consider the option to do the workshop in a hybrid mode; please specify when registering if you would be interested by this opportunity. All applications should be carefully motivated, and all participants on-site will be required to present their work as a contributed talk or poster (please indicate if you have any preference). We plan on having a few slots for contributed oral presentations and a poster session.