Machine learning interatomic potentials (MLIPs) allow quantum-mechanical accuracy to be achieved at a fraction of the cost of direct DFT calculations, enabling simulations at experimentally relevant timescales and system sizes. My work focuses on training and fine-tuning equivariant GNN potentials for complex inorganic and molecular systems.

Classical molecular dynamics treats nuclei as point particles, neglecting quantum effects such as zero-point energy and tunneling that are particularly important for light nuclei and at low temperatures. Path integral molecular dynamics (PIMD) addresses this by representing each nucleus as a ring polymer of classical beads, recovering quantum statistical mechanics exactly in the limit of many beads. My work combines PIMD with ML interatomic potentials to make quantum nuclear simulations tractable for large systems.

NMR spectroscopy is one of the most powerful probes of local structure and dynamics in materials, but first-principles computation of NMR parameters is expensive, limiting achievable statistical sampling. I develop NMR-ML models that predict NMR parameters for various nuclei at a fraction of the computational cost. A key focus is xenon NMR in porous liquids and carbon nanoporous materials.

This part of my research goes back to my PhD period and it deals with the precise structural characterization of disordered inorganic oxyfluorides, making conventional diffraction techniques insufficient for full structural resolution. By combining multi-nuclear solid-state NMR spectroscopy of various nuclei, including ¹⁹F, ¹H, ²³Na, ⁸⁷Rb, and ⁹³Nb, powder X-ray diffraction, and DFT calculations, I build and validate structural models that capture both the average and the short-range crystal structures. This NMR crystallography strategy has been applied to transition metal oxyfluorides, revealing preferential anion ordering, correlated disorder, and second-order Jahn-Teller distortions that govern the physicochemical properties of these materials.

Deploying ML potentials at scale requires careful integration with HPC infrastructure. I develop and document production-ready workflows for running packages on both AMD ROCm (LUMI) and NVIDIA CUDA (Mahti and Puhti) architectures, resolving compatibility issues between ML libraries, MPI implementations, and GPU backends. These workflows are made publicly available.