YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
HIP: Hessian Interatomic Potentials Without Derivatives
Available checkpoints
hip_v3-1(recommended): same weights aship_v3, with fixed graph settings stored in the checkpoint (6 Ã… Hessian graph,max_neighbors1000).hip_v2-1: same weights aship_v2, with fixed graph settings (5 Ã… Hessian graph,max_neighbors1000).hip_v3: Trained end-to-end with Muon, direct forces.hip_v2: Used for the paper. Trained end-to-endhesspred_v1: Used for the paper. Finetuned from HORM EquiformerV2 to predict Hessianship_mace: Unpublished. Used for https://github.com/BurgerAndreas/hip-mace
Graph settings for inference
hip_v2 and hip_v3 were trained with a 5 Ã… backbone graph, and the Hessian head saw atom pairs only up
to 5-12 Ã… (largest pair distance in the training set: 12.21 Ã…). A fully connected Hessian graph at
inference (the old code default for these checkpoints) makes Hessians of larger molecules much worse
(40x on PubChem molecules with 30-100 atoms). Use a 5-7 Ã… Hessian graph and no neighbour cap:
hip_v3-1.ckpt and hip_v2-1.ckpt carry these settings. They need the HIP code from
https://github.com/BurgerAndreas/hip main after 2026-11-11 (hessian_graph_cutoff); with that code the
original hip_v2.ckpt/hip_v3.ckpt also fall back to a 5 Ã… Hessian graph.
Benchmarks: https://github.com/BurgerAndreas/hip/tree/natcompsci-rebuttal/results_cutoff_ablation
| file | sha256 |
|---|---|
ckpt/hip_v3-1.ckpt |
040126d64c134c56e93839fa334853a42eb5a3b3b4b9de5db6ede5b66e2364e9 |
ckpt/hip_v2-1.ckpt |
6beda9642a018b37fa743335f8654297dc7c0ca1ae555a311cb7826b55b21843 |
For new models, the ideal settings going forward are a 5 Ã… backbone and a 5 Ã… Hessian graph with no
neighbour cap and a smooth edge envelope (configs/train_envelope.yaml in the code repo).
Citation
@misc{burger2025hiphessian,
title={Shoot from the HIP: Hessian Interatomic Potentials without derivatives},
author={Andreas Burger and Luca Thiede and Nikolaj Rønne and Varinia Bernales and Nandita Vijaykumar and Tejs Vegge and Arghya Bhowmik and Alan Aspuru-Guzik},
year={2025},
eprint={2509.21624},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2509.21624},
}