mofchecker-next

A fast, drop-in replacement for MOFChecker 2.0 — same diagnostics, same API, ~12× faster on the geometric diagnostics (~7× with open-metal-site detection), built on Rust kernels and rustworkx graph algorithms.

Designed for the workloads where the original is painful: validating thousands of model-generated MOFs (e.g. from a diffusion model), where the slow paths — floating-solvent extraction and dimensionality — dominate.

💪 Getting started

from mofchecker_next import MOFChecker

mc = MOFChecker.from_cif("structure.cif")     # also .from_ase(atoms) / MOFChecker(structure)
mc.has_atomic_overlaps, mc.has_lone_molecule, mc.has_oms, mc.metal_number
descriptors = mc.get_mof_descriptors()        # OrderedDict of every diagnostic

Validating many structures? mofchecker_next.batch parallelizes across structures, builds each graph once, and never aborts the run on a single bad structure:

from mofchecker_next.batch import check_structures

# inputs may be CIF paths, pymatgen Structures, ASE Atoms, or a mix
results = check_structures(inputs, n_workers=16)          # all CPUs by default
bad = [r for r in results if r["has_atomic_overlaps"]]

# subset to skip work: composition-only descriptors skip graph construction entirely
fast = check_structures(inputs, descriptors=["has_atomic_overlaps", "has_overcoordinated_c"])

Each result is a dict with index, id, n_atoms, and the requested descriptors. DEFAULT_DESCRIPTORS is the in-scope diagnostic suite (including bit-exact has_high_charges); ALL_DESCRIPTORS adds metadata, symmetry, and graph hashes. A structure that fails gets an error field (on_error="record") instead of aborting the batch.

🚀 Installation

pip install mofchecker-next

Latest from source (needs a Rust toolchain):

pip install git+https://github.com/henk789/mofchecker-next.git

✨ Why use it

• ⚡ Fast. On 150-atom QMOF MOFs, the geometric diagnostic set runs in ~101 ms/structure single-core vs ~1.20 s for MOFChecker 2.0 (~12×); ~77 structures/s across 10 cores. The win is the floating-solvent and 3D-connectivity paths, ported off networkx onto rustworkx; the numeric kernels (distances, contacts, connected components, OMS Voronoi/order-parameters, EQeq) are Rust.
• 🔌 Drop-in. MOFChecker-compatible class — same properties, same get_mof_descriptors(). Switch the import and existing code keeps working.
• ✅ Parity-verified. 100% agreement with MOFChecker 2.0 on real QMOFs (4500/4500 descriptor-comparisons over 250 structures × 18 descriptors; 16/16 on the reference test CIFs). See Parity below.
• 📦 Built for batches. Parallel check_structures, graph built once per structure, failures isolated.
• 🔋 Bit-exact charges. has_high_charges is a faithful Rust port of EQeq (bit-exact equilibrated charges).
• 🔁 Reproducible. symmetry_hash is deterministic (the reference's depends on Python hash randomization).

⚡ Performance

30 QMOF relaxed 150-atom MOFs, cold batch (each structure processed once), MOFChecker 2.0 vs mofchecker-next on identical inputs, dedicated compute node, 1 vs 10 cores. Reproduce with scripts/benchmark_throughput.py (the structure set is pinned + fingerprinted so before/after runs are comparable).

Geometric diagnostic set (the structural/graph checks; excludes open-metal-site detection):

Full geometric suite (includes has_oms open-metal-site order parameters):

Where the speedup comes from — it is concentrated in one check. MOFChecker's floating-solvent detection (has_lone_molecule) builds a 3×3×3 supercell graph via pymatgen StructureGraph.__mul__ (networkx union/relabel of 27 copies, ~940 ms/structure here) and 3D-connectivity runs Larsen dimensionality over networkx. These are replaced by direct integer image-offset algorithms on a rustworkx graph — O(N+E), no supercell — making has_lone_molecule ~94× faster (~940 ms → ~10 ms); it dominates the reference's runtime, so it drives the speedup. has_oms now uses Rust for Voronoi facet-neighbor selection and pymatgen-compatible local order-parameter formulas. Speedup also grows with structure size and is higher on generated/distorted structures (more disconnected fragments).

⚙️ How it works

Python owns CIF/structure loading, pymatgen integration, and orchestration. The heavy lifting is delegated:

• Rust (_rust PyO3 extension): minimum-image distances, short contacts, neighbor candidates, connected components, graph degrees, OMS Voronoi/order-parameters, and the EQeq charge solve.
• rustworkx (checks/_subgraph_rx.py): floating-solvent / lone-molecule detection (finite connected components via an image-offset consistency test) and Larsen dimensionality (rank of the lattice-image vectors a component spans). These replace the networkx-heavy paths.
• structuregraph_helpers is retained for the logic-critical, non-hot pieces it does well: graph construction (tuned VESTA cutoffs) and the Weisfeiler–Lehman graph hashes.

The structure graph is built once per MOFChecker and reused across all checks.

✅ Parity

Verified against a MOFChecker 2.0 checkout (used only as a behavioral oracle) via the harnesses in scripts/:

• Real QMOFs: 4500/4500 descriptor-comparisons (250 structures × 18 descriptors) — 100%.
• Reference test CIFs: 16/16.
• Generated (distorted) structures: 3899/3900. The single difference is has_lone_molecule, where mofchecker-next is more correct — see Limitations.

Reproduce: scripts/qmof_parity.py (real QMOFs), scripts/generated_parity.py (generated CIFs), scripts/validate_subgraph_rx.py (floating-solvent port). Point them at a local QMOF CIF directory with QMOF_DIR=....

⚠️ Limitations & deliberate differences

• Healing not implemented. adding_hydrogen / adding_linker raise NotImplementedError.
• No porosity. is_porous returns None (no bundled Zeo++).
• has_lone_molecule is more correct than the reference. MOFChecker 2.0's supercell + in-cell-filter heuristic silently misses finite molecules that wrap the unit-cell boundary (the origin-cell copy is truncated at the supercell face). mofchecker-next detects them via a topological finite-component test. This is the only descriptor that ever disagrees with the reference, only on pathological/distorted structures (0 disagreements on real QMOFs).
• Graph construction is still the floor. pymatgen's VESTA neighbor-finding is unchanged; the speedup is in the graph algorithms, not bond perception.
• Determinism. symmetry_hash is deterministic by design and will not match the reference's randomized value across runs.

⚖️ License

The published package is GPLv2, because it bundles py/mofchecker_next/eqeq/ — a faithful translation of EQeq (GPLv2, see py/mofchecker_next/eqeq/LICENSE) — and the GPL governs the combined work. The non-eqeq sources are MIT (LICENSE); for an MIT-only build, omit the eqeq subpackage and the has_high_charges diagnostic.

The MOFChecker 2.0 checkout used as the behavioral oracle (ANCSA 1.0) is not redistributed; see external/REFERENCE.md to reproduce it locally.

🛠️ For developers

python -m maturin develop --release          # build the Rust extension into the venv
python -m pytest -q                          # Python tests
cargo test --release --manifest-path rust/Cargo.toml   # Rust tests
# Source builds need libclang available for qhull-sys/bindgen.

git tag v0.1.0 && git push origin v0.1.0