Ponentes
Descripción
Simulating molecular systems across chemically relevant time and length scales remains a central challenge in computational chemistry. First-principles methods such as Hartree–Fock and density functional theory provide accurate references, but become prohibitively expensive beyond modest system sizes. At the same time, generating the large-scale datasets required for modern machine-learned force fields (MLFFs) — including efforts such as Meta FAIR’s OMol25 — is itself an HPC-intensive task based on massive first-principles calculations. MLFFs bridge the gap between accuracy and tractability by combining reference data across levels of theory into transferable models, evolving from kernel methods to modern equivariant graph neural networks.1 In this talk, we present SO3LR, a general-purpose MLFF implemented in JAX that exemplifies this new generation of molecular AI.
Beyond data generation, training modern MLFFs requires large-scale GPU computing, efficient distributed workflows, and software frameworks capable of exploiting accelerator architectures. Using the Luxembourgish ecosystem around MeluXina as a case study, we discuss how software-stack choices such as JAX/JAX-MD versus PyTorch affect compilation, differentiability, vectorization, and scalability, ultimately shaping feasible molecular dynamics and training workflows on modern supercomputers.
We also present user-friendly workflows built around MARS, MARS-ROVER, and MARS-MS, a family of tools designed to make MLFF-based simulations accessible for chemical applications. These packages automate con-
formational exploration, prediction of infrared and mass spectrometry signatures, and reaction-pathway searches, enabling pretrained MLFFs to serve not only as fast molecular dynamics engines but also as practical platforms for interpreting experimental observables. Applications include organic molecules and peptides, SN2 reaction pathways, IR and mass spectroscopy, and conformational ensemble analysis. More broadly, pretrained MLFFs are reshaping atomistic simulation by dramatically reducing inference-time costs, enabling accurate simulations on single-GPU workstations while extending quantum-informed modeling toward biomolecular systems. As molecular AI increasingly targets GPU- and TPU-based platforms, balancing mixed-precision performance with
the numerical reliability required by scientific simulations will become a defining challenge.