Step 3: Validator¶

The Validator takes TS guess geometries from path guessers and validates them with quantum chemistry calculations. A TS is accepted if it has one imaginary frequency and the IRC check matches reaction connectivity.

Validation process¶

flowchart LR
    TS[TS Guess] --> OPT[Geometry<br/>Optimization]
    OPT --> FREQ[Frequency<br/>Calculation]
    FREQ --> CHECK{Single imaginary<br/>frequency?}
    CHECK -->|Yes| IRC[IRC Calculation]
    CHECK -->|No| REJECT[Rejected]
    IRC --> VERIFY{Connects R & P?}
    VERIFY -->|Yes| VALID[Validated TS]
    VERIFY -->|No| REJECT

Available validators¶

GFN2-xTB validator¶

Fast semi-empirical validation.

python -m motsart.validator.base_validator env=test validator_cfg=test validator=xtb env.rxn_num=0

DFT validator (ORCA)¶

More accurate but slower.

python -m motsart.validator.base_validator env=local validator_cfg=local validator=dft env.rxn_num=0

Note

ORCA must be installed and configured in env.orca_path.

MLIP validator (OMol25, via ORCA ExtOpt)¶

Drives ORCA's TS optimizer with a FAIRChem OMol25 machine-learned interatomic potential (default: eSEN-sm-conserving) through ORCA's otool_external (! ExtOpt) interface. Energies and gradients come from the MLIP while ORCA's own OptTS/IRC/NumFreq drive the geometry, so the engine is swappable with xtb/dft and nothing else changes.

python -m motsart.validator.base_validator env=local validator_cfg=local validator=mlip env.rxn_num=0

Note

Requires fairchem-core + torch in the environment, and access to the eSEN checkpoint (HuggingFace-gated registry name esen-sm-conserving-all-omol, or a local *.pt via MOTSART_MLIP_MODEL). The device is auto-selected (CUDA if available, else CPU). The MLIP runs gas-phase (no implicit solvent). The external-tool wrapper lives at src/motsart/validator/orca_validator/orca_external_tools/mlip_external.py; see experiments/ for engine-comparison run/analysis scripts and a no-GPU plumbing test.

Configuration¶

Validator config (`validator_cfg`)¶

Parameter	Description
`SP_maxcore`	Maximum memory per core (MB) for single-point calculations
`SP_nprocs`	Number of parallel processes
`SP_MaxIter`	Maximum optimization iterations
`IRC_MaxIter`	Maximum IRC iterations
`skip_full_irc`	If `true`, use heuristic IRC via graphRC only
`path_guessers_to_validate`	List of TS methods to validate
`require_no_extra_bonds`	graphRC strictness. If `false` (default), a TS passes when all expected bond changes are recovered, tolerating extra (unexpected) ones; if `true`, any extra bond change also rejects the TS

Computing statistics¶

When you ran the pipeline, e.g. in parallel on a cluster for hundreds of reactions, you want to check statistics on success metrics of SP optimization or IRC validation. The custom script below was used to compare two different runs: guesses generated with GoFlow (TsOptNet in the paper) and those from another path guesser, such as RMSD-PP. Please adjust it to your liking.

python -m motsart.validator.compute_stats \
  --cluster-folder /data/results \
  --learning-folder /home/lgalustian/projects/motsart/results_goflow \
  --validator DFTValidator \
  --output-csv /home/lgalustian/projects/motsart/results_goflow/stats_al.csv \
  --cluster-ts-method racer_ts \
  --al-ts-method learning \
  --mode both

Output¶

Results are saved to results*/R{rxn_id}/validation/{method}/:

optimized TS geometries
frequency outputs
IRC trajectories
per-guess validation status CSV