Predicting ¹²⁹Xe NMR chemical shifts in host–guest systems from first principles remains computationally prohibitive at the scale required for meaningful statistical sampling. Here we develop equivariant graph neural networks trained on quantum-chemical shielding data to learn the relationship between local host–guest geometry and xenon chemical shift in porous liquid systems. The model is applied to CC3 organic cages in solution, revealing how cage breathing, solvent reorganization, and xenon siting collectively shape the observed NMR signal. Our approach reduces computational cost by several orders of magnitude relative to DFT-based sampling while reproducing experimental trends quantitatively. The results demonstrate that equivariant architectures can capture the subtle many-body interactions governing NMR observables in soft porous hosts, opening a route to routine ML-accelerated NMR prediction in complex guest–host materials.

Phase transitions in transition metal phosphides often involve subtle structural reorganizations that are difficult to resolve experimentally. Here we fine-tune the MACE-MH-1 atomistic foundation model on density functional theory data for ruthenium phosphide (RuP) and deploy it in large-scale molecular dynamics simulations spanning temperatures from 100 to 500 K. Our simulations reveal a two-stage phase transition: a first transition near 180 K associated with local symmetry breaking and the onset of Ru trimerization, and a second transition near 330 K marking the full recovery of the high-symmetry Pnma structure. Radial distribution functions, structure factors, and space-group analysis consistently identify an intermediate phase of reduced symmetry between the two transitions. These findings provide atomistic insight into the structural chemistry of RuP and demonstrate the power of fine-tuned foundation models for uncovering complex phase behavior in functional inorganic materials.

Accurate prediction of NMR chemical shifts across multiple nuclear species simultaneously is a key challenge for computational materials chemistry. In this work we develop and benchmark a family of machine learning NMR models — including kernel ridge regression, SchNet, and MatTen architectures — trained on relativistic DFT shielding data for multi-element systems. We systematically assess the role of descriptor choice, training set size, and architecture depth on the transferability of learned shielding models across chemical environments. The models are applied to predict ¹H, ¹³C, and ¹⁵N chemical shifts in organic and hybrid inorganic–organic materials, achieving mean absolute errors competitive with state-of-the-art approaches at a fraction of the computational cost. Our results highlight the importance of equivariant representations for capturing the angular dependence of shielding tensors, and provide practical guidelines for building multi-element ML NMR models in heterogeneous material systems.