Runnable notebooks. Each one is self-contained and uses the files in
files/. They are also rendered in the
documentation, which builds this directory
directly — so a notebook fixed here is fixed there.
Most of these call a neural network potential. The first run downloads the
AIMNet2 weights into ~/.cache/aimnet (a few hundred MB, once per model), so
give the first cell a moment. Pass use_gpu=False if you have no CUDA device —
Auto3D treats a requested-but-missing GPU as a fatal error, not a fallback.
| Notebook | What it covers |
|---|---|
| quickstart | Smallest useful example: SMILES in, ranked conformers out |
| tutorial | The same ground at a slower pace, with the output explained |
| performance_tuning | Batch size, memory, GPU selection, what actually moves the needle |
| large_scale_processing | Chunking and running a library that will not fit in memory |
Each wraps one Python API function, and each has a matching CLI command.
| Notebook | API | CLI |
|---|---|---|
| single_point_energy | calc_spe |
auto3d energy |
| geometry_optimization | opt_geometry |
auto3d optimize |
| thermodynamic_calculation | calc_thermo |
auto3d thermo |
| tautomer | get_stable_tautomers |
auto3d tautomers |
calc_thermo needs the ase extra: pip install "Auto3D[ase]".
| Notebook | What it covers |
|---|---|
| virtual_screening | Preparing a screening library |
| tautomer_protomer_analysis | Tautomer and protomer states for drug-like molecules |
| stereochemistry | Enumerating and keeping track of stereoisomers |
| docking_integration | Handing conformers to a docking program |
| Notebook | What it covers |
|---|---|
| reaction_thermodynamics | Reaction energies from optimized conformers |
| boltzmann_populations | Populations and conformational averaging |
| strain_energy | Ligand strain |
| molecular_descriptors | 3D descriptors for ML/QSAR |
| Notebook | What it covers |
|---|---|
| md_preparation | Preparing structures for molecular dynamics |
| qm_refinement | Refining Auto3D output with a QM program |
| using_custom_NNP | Wrapping your own potential to the CustomNNP contract |
| tautomer_with_userNNP | The same, driving tautomer ranking |
Writing a custom potential is the one place where getting the contract slightly
wrong fails in a confusing way. Read
the custom NNP guide
alongside those last two: your model implements
forward(species, coords, charges) -> energies (species first, energies
only), which is the mirror image of Auto3D's internal adapter interface.