DreamLufei/invDesMobility-megnet-formation-energy

InvDesMobility MEGNET Formation-Energy Screening Model

This repository contains the MEGNET/formation-energy surrogate used as one of the screening stages in the InvDesMobility candidate filtering pipeline.

Paper and Repositories

• —Paper: InvDesMobility: a reliability-gated first-principles feedback framework for closed-loop materials discovery
• —arXiv: https://arxiv.org/abs/2606.16133
• —Project page: https://dreamlufei.github.io/invDesMobility/
• —GitHub: https://github.com/DreamLufei/invDesMobility
• —Loop repository: https://github.com/DreamLufei/invdesmobility_loop
• —Mobility workflow: https://github.com/DreamLufei/2d-mobility
• —Zenodo: https://doi.org/10.5281/zenodo.20475023

Files

• —FormEGNN-weight.hdf5: model weights.
• —megnet_formation_energy.hdf5.json: model/config metadata.

Intended Use

Use this model only as a surrogate formation-energy screen before downstream DFT validation. It is not a substitute for first-principles total-energy calculations.

Training Data

This model is released as the fitted formation-energy surrogate used by the InvDesMobility screening pipeline. The model repository contains the exact weights and MEGNet graph-construction metadata used for inference. The raw training records for this surrogate are not packaged in this model repository.

The companion DreamLufei/invDesMobility-data dataset repository contains the InvDesMobility generated pools, feedback records, and screening data used by the closed-loop workflow, but it should not be interpreted as the complete standalone training set for this MEGNet surrogate.

Model and Training Configuration

The sidecar JSON records the inference-time MEGNet graph converter:

• —CrystalGraph converter from megnet.data.crystal.
• —MinimumDistanceNNAll neighbor strategy with cutoff 5.
• —DummyScaler target scaler.

Optimizer history, training loss curves, and a standalone training command are not present in the archived sidecar. The repository is therefore intended to reproduce the screening predictions used by the project, not to fully retrain the MEGNet surrogate from scratch.

Evaluation

No independent held-out benchmark table for this MEGNet surrogate is included in this artifact archive. In InvDesMobility it was used as an intermediate filter before ALIGNN ranking and deterministic DFT validation. Candidate retention decisions and formation-energy thresholds are recorded in the workflow logs and generated screening CSV files distributed with the companion data archive.

Limitations

The model is a screening aid. Generated candidates retained by this stage still require subsequent electronic, structural, and first-principles mobility validation. Formation-energy predictions from this surrogate should not be reported as final thermodynamic stability evidence.