gcnwm/fyp-active-learning-models

FYP raw and compiled active-learning checkpoints for rMD17 ethanol

This repository contains fresh model publication artifacts from the bachelor-thesis active-learning study on rMD17 ethanol. Each bundle corresponds to the best final-iteration seed by forces MAE for one architecture/strategy family.

The publication was rebuilt from scratch from the local results/al tree rather than reusing previously uploaded assets.

Included artifact types

• Raw MACE checkpoints: original .model files copied from the final selected iteration.
• Compiled MACE checkpoints: official LAMMPS exports produced with python -m mace.cli.create_lammps_model --format libtorch.
• Raw NequIP checkpoints: original Lightning .ckpt files copied from the selected iteration.
• Packaged NequIP checkpoints: official .nequip.zip packages produced with nequip-package build.
• Compiled NequIP checkpoints: official AOTInductor artifacts produced with nequip-compile --mode aotinductor --device cuda --target ase.

Selected bundle summary

Bundle	Architecture	Strategy	Best seed	Final iteration	Forces MAE (meV/Å)	Energy MAE (meV)	Raw files	Packaged files	Compiled files
mace-random	MACE	RANDOM	2	9	9.20	1.71	1	0	1
mace-qbc	MACE	QBC	2	9	8.51	1.60	4	0	4
mace-mhc	MACE	MHC	2	9	9.97	2.07	1	0	4
nequip-random	NequIP	RANDOM	3	9	8.89	4.74	1	1	1
nequip-qbc	NequIP	QBC	1	9	7.57	1.48	3	3	3

Important details

• mace-qbc/ contains the full four-member committee for the selected QBC seed.
• mace-mhc/compiled/ contains one compiled LAMMPS export per available model head.
• nequip-qbc/ contains one raw/package/compiled triplet per selected committee member.
• The NequIP training checkpoints embed legacy absolute workspace paths, so packaging was performed from a temporary local mirror with rewritten checkpoint-path prefixes before running the official NequIP CLI.
• The compiled NequIP .nequip.pt2 artifacts are device-specific AOTInductor exports built for cuda on this machine; the .nequip.zip packages are the more portable redistribution artifact.

Provenance

• Hugging Face repo: https://huggingface.co/gcnwm/fyp-active-learning-models
• Source thesis workspace commit: 8fea4f53dcdf581ca23a820ab1cc8de9708f7aac
• Companion production-code repository: https://github.com/gcnwm/fyp-production
• Companion production-code commit: 433f9ee3872a84cfeabf19a5d64fca9277d5cf09

Citation

If you use these checkpoints, please cite the accompanying thesis project and the upstream MACE, NequIP, and rMD17 papers.