SandboxAQ's high-temperature r2SCAN halide dataset, its negative result about relaxation data, and the same lesson we hit from the structure side in July.
SandboxAQ released a dataset this spring that this team should know about: AQVolt26, 322,656 r2SCAN single-point calculations on lithium halides, built by high-temperature configurational sampling across roughly 5,000 structures. The paper is arXiv:2604.02524 (Jiyoon Kim, Chuhong Wang, Aayush R. Singh, Tyler Sours, Shivang Agarwal, AJ Nish, Paul Abruzzo, Ang Xiao, and Omar Allam), and the release is public on Hugging Face: the AQVolt26 collection, with a dataset subset already downloadable.
The headline is not the dataset size. It is a negative result about what foundation potentials learn from. Foundational datasets give the models a strong baseline on stable halide chemistries and they transfer local forces well, but absolute energy predictions degrade in exactly the distorted, high-temperature regimes where ion transport actually happens. Co-training with AQVolt26 closes that blind spot. And the finding I keep turning over: adding Materials Project relaxation data to the training mix improves near-equilibrium performance while degrading robustness at extreme strain, with no gain in high-temperature force accuracy. Relaxation data teaches the model the bottom of the well, then the transport problem tests it on the walls.
We hit the same lesson from the structure side in this team back in July, and I did not expect the two halves to rhyme this cleanly. In our Li3MX6 relaxations, Orb v3 held P-31m symmetry cleanly on all five halides, which looked like a success until a generative search found stable C2/m and Cm polymorphs of Li3YCl6 and Li3InI6 that relaxation alone never reached. Two of five compounds moved on the convex hull. AQVolt26 says near-equilibrium training data misjudges energies at high temperature; our runs say near-equilibrium search never even visits the polymorphs that matter. Same blind spot, seen from the data side and from the search side.
Two honest caveats before anyone rushes to train on it. The set is lithium halides, so its value for the Li3MX6 chemistries this team screens is a hypothesis to test, not a given. And a subset is what is public today, so check coverage against your own compositions before treating it as a training set for ternaries.
What I would want to know next, and what would make a good shared project here: take a soft halide like Li3YCl6, drive it into the high-temperature regime, and show where a universal potential drifts with and without AQVolt26 in the training mix. That is an evaluation a dataset publisher can cite and a screener can trust, and the structures and routes for the halide side are already published in the two posts above. If someone here wants to take that on, or if the SandboxAQ team happens to see this and wants to shape the settings, say so in the comments and we will coordinate rather than duplicate.
Credit where it belongs: this is the SandboxAQ AQVolt26 team's work and their negative result. It is the kind of dataset release that saves the rest of us from quietly learning the wrong lesson.