Abstract: Automated material discovery relies on the ability to accurately characterize synthesized materials, a task where solid-state nuclear magnetic resonance (NMR) spectroscopy plays a crucial role due to its atom-level insights. However, interpreting solid-state NMR spectra is challenging, often requiring quantum mechanical calculations that traditionally assume static materials. This assumption neglects the dynamic nature of materials at operational temperatures, leading to discrepancies between computational and experimental results. To overcome this limitation, we propose an approach that incorporates both molecular dynamics and transition state searching to model dynamic effects across timescales in NMR spectra using machine learning force fields (MLFFs). By fine-tuning atomistic foundation models, we achieve accurate, cost-effective MLFFs in an automated manner. We validate our approach through $^{17}$O NMR experiments on porous materials for carbon capture. Specifically we look at the metal-organic framework MFU-4l, demonstrating that existing methods for predicting spectra fail to match experimental observations, while our method achieves strong agreement. This workflow not only facilitates automated characterization of materials critical for carbon capture but also highlights the potential of ML-driven simulations in predicting material properties.