Dataset for: Predicting the Thermodynamic Stability of Ce/Ti Nanoclusters under Synthetic Conditions

Data record supporting the manuscript Predicting the Thermodynamic Stability of Ce/Ti Nanoclusters under Synthetic Conditions. Liping Liu and Haley Wellman contributed equally; Omar K. Farha and Laura Gagliardi are the corresponding authors. The study asks why two hexanuclear Ce/Ti clusters form under different synthetic conditions: a yellow CeIV4Ti2 cluster that crystallises under mild conditions, and a red CeIII2Ti4 cluster, reported for the first time in this work, that requires higher temperature, longer reaction times and a higher Ti precursor concentration. Answering it means distinguishing a thermodynamically disfavoured product from a kinetically hindered one, which requires sampling cluster configurations in explicit solvent rather than relaxing a single crystallographic geometry. The record provides the three computational ingredients of that workflow. First, a fine-tuned machine-learned interatomic potential: UMA-s was fine-tuned on 228 DFT-labelled Ce/Ti cluster configurations, augmented with 1000 OMol structures to retain the base model's general chemistry (1228 training / 25 validation structures with energies and forces). Included are the fine-tuning dataset as ASE-DB LMDB shards, the complete Hydra configuration and training log, and the production checkpoint, which reaches a validation energy MAE of 0.6 meV/atom and a force MAE of 7.5 meV/Å over 15 epochs. Second, the DFT benchmark that validates it: relaxed energies for three clusters — CeIV4Ti2, CeIII2Ti4, and the hypothetical CeIV2Ti4 used to decompose the stability difference — against pretrained UMA-s, pretrained UMA-m and the fine-tuned model. Fine-tuning reduces the per-atom energy error from 5.8–13.9 meV (UMA-s) and 2.2–6.2 meV (UMA-m) to 0.1–1.0 meV, the accuracy needed for accurate configurational sampling in explicit solvation. Third, the explicit-solvent configurational sampling: simulated-annealing molecular dynamics driven by the fine-tuned potential, for the CeIV4Ti2 cluster and the solvated Ce ion in explicit acetonitrile droplets. Annealing locates configurations roughly 1 eV below those obtained by direct relaxation of the crystal structures — the step that changes the predicted outcome. Together these support the paper's central result: despite requiring more forcing synthetic conditions, CeIII2Ti4 is the thermodynamic product, more stable than CeIV4Ti2 by 4.48 eV, and its formation only under more forcing conditions is consistent with kinetic limitation. Zenodo records are flat, so the bulk of this deposit ships as tar archives whose members carry paths relative to the bundle root. See ZENODO_CONTENTS.md for the archive map, per-directory contents, checksums, and extraction instructions. Note that the analysis scripts and Hydra configs are shipped unmodified and therefore contain absolute local paths that must be edited before re-running. The fine-tuned checkpoint is a derivative of Meta FAIR Chemistry UMA-s and is governed by the FAIR Chemistry License, not by this record's CC-BY-4.0 licence.