Quick reference

Page	What you do	Runs on	Key command or script
A1 Batched relaxation for screening	batched relaxation with TorchSim	NVIDIA GPU (Leonardo); CPU smoke test	pixi run smoke, scripts/test-leonardo-torchsim.sbatch
A2 Orb-v3, MatGL and dispersion	Orb-v3, MatGL and D3 on graphite	CPU or NVIDIA GPU	pixi run graphite-cpu, scripts/test-leonardo-orb.sbatch
A3 MatGL: relaxation, MD, lattice benchmark and training	MatGL relaxation, MD, benchmark, training	LUMI, one MI250X GCD	python examples/matgl/cli.py prefetch, scripts/lumi-matgl.sbatch
A4 Train and fine-tune a potential	fine-tune TensorNet from PBE to r2SCAN	LUMI	scripts/lumi-training-setup.sh, scripts/lumi-training.sbatch
A5 NEB: migration barriers with MLIPs	CI-NEB of a Li hop in LiFePO4	LUMI	python examples/neb/cli.py prefetch, scripts/lumi-neb.sbatch
A6 One model for materials and molecules: Orb and UMA	Orb-v3, OrbMol and UMA with charge and spin	LUMI	python examples/uma/cli.py prefetch --model orb (or uma), scripts/lumi-uma.sbatch
B2 Build the ALCHEMI image	build the ALCHEMI image	GH200 site	scripts/build-alchemi-sif.sh
B3 Build native LAMMPS with ML-IAP, Kokkos, and MPI	build LAMMPS with ML-IAP and Kokkos	GH200 site	scripts/build-lammps-mpi.sbatch
B4 Simulate silicon with MACE to B6 Run the same silicon case with LAMMPS	silicon MD: one trajectory, a batch, LAMMPS replicas	one GH200	scripts/run-alchemi.sh, scripts/run-lammps.sh
B7 Read a completed shared-GPU benchmark	read a finished shared-GPU benchmark	no GPU	none
B8 One silicon simulation on more than one GPU	one simulation on 1, 2 and 4 GPUs	GH200 node	scripts/lammps-mpi-size-benchmark.sbatch

Submit Slurm jobs with sbatch --account=<PROJECT> <script>; site set-up is in Before you start, variables in Environment variables. UMA weights are gated: accept the licence on the Hugging Face model page and use a token. Every page shows saved results, so it can be read without a GPU.