Published November 7, 2026 | Version 1.0.0

LEH-MACE: code and trained models for a local excitation Hamiltonian for excited-state machine-learned potentials

Authors/Creators

  • 1. Karlsruhe Institute of Technology

Description

Code and trained models of the paper Toward universal excited-state machine-learned potentials: a local excitation Hamiltonian for size consistency and transferability (P. Friederich).

LEH-MACE-v1.0.0.zip: snapshot of https://github.com/aimat-lab/LEH (MIT license): the exmlip package (LEH and direct multistate readouts on a MACE encoder, training, ASE calculator, batched excited-state dynamics, TURBOMOLE interface), the configuration and summary of every model of the paper, experiment, analysis, and figure scripts with their JSON results, cluster scripts, and the export and loader of the dataset.

leh_mace_checkpoints.tar: checkpoints of all 152 trained models of the paper (all seeds): LEH-MACE and direct readouts (best.pt, meta.pt), X-MACE (xmace.model), MS-ANI (model.pt), and fine-tuned OMNI-P2x ensembles (member1-3.pt). Unpack it in the root of the repository; every run directory then contains its configuration and checkpoint. MODELS.csv lists every model with its architecture, training data and split, and seed; the default production models are experiments/al/runs/prodw_leh_v4_s0 (LEH-MACE) and experiments/al/runs/prodw_softmin_v4_s3 (Direct-softmin).

The dataset is a separate record (CC BY 4.0, https://doi.org/10.5281/zenodo.24042143).

Files

LEH-MACE-v1.0.0.zip

Files (1.0 GB)

Name Size
md5:f0905eb68f13ebe5e535e09ae08ab713
15.4 MB Preview Download
md5:45884aa20db97cdf3305ceaef71eeb47
1.0 GB Download
md5:93dd630ed8b8fb9d011a7b5cff2251e9
19.1 kB Preview Download

Additional details

Related works

Is derived from
Dataset: 10.5281/zenodo.24042143 (DOI)
Is supplement to
Software: https://github.com/aimat-lab/LEH (URL)