Examples#

Every example notebook in the repository’s examples/ directory, rendered with its executed outputs (training curves, parity plots, MD observables, and the fidelity tables) so you can read them without running anything. To run one yourself, see Run the Example Notebooks for the required environments and kernels; the notebooks below are the committed, fully executed versions.

Data#

Data

ANI (dnn)#

Training ANI from scratch on rMD17, then the four published training sets (ANI-1, ANI-1x, the coupled-cluster ANI-1ccx with its transfer-learning recipe, and the seven-element ANI-2x set), each paired with its matching model preset.

ANI (dnn)

PhysNet (dnn)#

PhysNet (dnn)

SchNet (cnn)#

Training the paper-architecture SchNet on its own MD17-style benchmark (rMD17 ethanol, energies + forces through the hub), then driving thermostat-free NVE dynamics with the trained model to demonstrate the paper’s energy-conservation-by-construction claim.

SchNet (cnn)

NequIP, MACE, Allegro, CACE (gnn)#

The shared Argon train/evaluate/deploy series (one pair of notebooks per model), plus a block-by-block walkthrough of the MACE architecture, and the MACE foundation models in action: mace_foundation_molecules.ipynb loads MACE-OFF23 with one MACE.from_foundation() call and runs the butane torsion profile against OPLS-AA, the water dimer against CCSD(T)/CBS, and a Trainer fine-tune to a new DFT reference (rMD17 malonaldehyde); mace_foundation_materials.ipynb screens equations of state (Si, Al, NaCl) across the MACE-MP generations (MP-0, MPA-0, OMAT-0).

NequIP, MACE, Allegro, CACE (gnn)

Long-range: Latent Ewald Summation (gnn)#

Long-range: Latent Ewald Summation (gnn)

Dispersion: DFT-D4 (common)#

The charge-dependent DFT-D4 dispersion correction as a model-agnostic add-on. d4_paper_examples.ipynb reproduces examples of the D4 paper (Caldeweyher et al. 2019): the charge-scaling function of fig 2, the charge- and CN-dependence of the carbon and hydrogen polarizabilities of fig 5, the atom-in-molecule polarizabilities and the molecular C6 coefficient of (3Z)-hexen-1-yne (fig 3b), the molecular C6 coefficients of the DOSD benchmark against the experimental dipole-oscillator-strength values (table III), and the D4 vs D3(BJ) dispersion contributions to the S22 interaction energies. d4_benchmark.ipynb benchmarks the xnn implementation against the reference dftd4 code (accuracy on S22 and crystals; timing on CPU and GPU versus system size, with the upstream and MD-friendly cutoffs) and shows D4 correcting a short-range MLIP through every deploy channel.

Dispersion: DFT-D4 (common)

Dispersion: DFT-D3 (common)#

The geometry-dependent DFT-D3 correction (Grimme et al. 2010; BJ damping Grimme, Ehrlich & Goerigk 2011) as the same model-agnostic add-on. d3_paper_examples.ipynb reproduces examples of the two papers: the rare-gas and carbon C6 coefficients of table II and the rare-gas C9 of table III (2010), the CN-dependent C6 curves of fig 5, the two-carbon dispersion energy of fig 1, the zero- vs BJ-damped argon dimer of fig 1 of the 2011 paper, the three-body share of the graphene bilayer binding (table VII), and the DOSD molecular C6 comparison of fig 6. d3_benchmark.ipynb benchmarks xnn against the reference s-dftd3 (S22, crystals, all damping functions; timing on CPU and GPU) and deploys a D3-corrected MLIP through every channel.

Dispersion: DFT-D3 (common)

BAMBOO (hybrid)#

BAMBOO (hybrid)

ReaxFF / ReaxFF-nn (ffnn)#

Training a reactive force field by gradient descent: a generic seed library is fit to rMD17 malonaldehyde energies and forces through the standard pipeline, then exported as a portable ffield.json. The companion notebook runs ASE molecular dynamics with the trained library and analyses the reactive descriptors – per-pair bond orders, geometry-dependent EEM charges, and a smooth bond-dissociation scan – including an honest look at what equilibrium-only training data cannot constrain.

ReaxFF / ReaxFF-nn (ffnn)

OPLS / L-OPLS (ffnn)#

The fixed-topology classical force field, validated against its own literature: opls_conformational_energetics.ipynb reproduces the relaxed torsional energies of Table 1 of Jorgensen et al. (1996) with the paper’s dihedral-driver protocol (ethane, propane, butane, methanol, ethanol), and opls_lopls_torsion_refit.ipynb first compares the hexane torsion profile of OPLS-AA and L-OPLS (Siu et al. 2012) and then re-derives the L-OPLS CT-CT-CT-CT torsion by gradient descent — mark dihedral_v trainable, fit conformer energies, recover the published Fourier coefficients to machine precision — before exporting the trained library and checking NVE energy conservation.

OPLS / L-OPLS (ffnn)

DREIDING (ffnn)#

The rule-generated generic force field, tested against the paper that defined it: dreiding_conformational_energetics.ipynb reproduces the single-bond rotational barriers of Table XI (fourteen molecules, mean difference from the paper’s own calculated column ~0.01 kcal/mol) and the butane and cyclohexane entries of Table XII, all from relaxed scans, and shows where DREIDING’s deliberate simplifications part company with experiment. dreiding_refit_aromatics.ipynb then treats the generators as trainable parameters: refit them on benzene alone against rMD17 PBE forces and ask what that does to naphthalene and toluene, neither of which was trained on — a direct test of the transferability DREIDING claims.

DREIDING (ffnn)

Deployment: MDI (common)#

Serving a trained checkpoint as a MolSSI Driver Interface engine: train on the hub argon data, launch xnn mdi, validate the wire protocol against direct evaluation, and drive NVE molecular dynamics from a minimal Python driver. The companion notebook then replaces the Python driver with LAMMPS (fix mdi/qm): same engine, production driver, with LAMMPS-side thermodynamics and a radial distribution function. The engine is model agnostic, so the same workflow serves any family’s checkpoint.

Deployment: MDI (common)

Fidelity checks#

Block-by-block numerical verification of each xnn implementation against its upstream reference (see Model Fidelity to Upstream Codes for the summary of what matches and to what precision). SchNet is the exception that proves the rule: a clean-room build verified against the manuscripts’ equations instead of a reference code.

Fidelity checks