Pipeline Overview¶

Multi-stage pipeline of moTSart for transition-state discovery. Each stage consumes artifacts from the previous stage.

Architecture¶

flowchart TD
    INPUT["Input Reactions (CSV)<br/><code>rxn_id, rxn_smiles</code>"]
    CF["<b>Step 1: Complex Finder</b><br/>Evolut. algorithm + AFIR<br/><code>motsart.complex_finder</code>"]
    PG["<b>Step 2: Path Guessers</b><br/>RMSD-PP then RacerTS<br/><code>motsart.path_guessers</code>"]
    VAL["<b>Step 3: Validator</b><br/>xTB or DFT + IRC<br/><code>motsart.validator</code>"]
    AL["<b>Optional: Learning</b><br/>Data prep + model TS evaluation<br/><code>motsart.learning</code>"]

    INPUT --> CF --> PG --> VAL --> AL
    AL -.->|improved TS guesses| PG

Main abstractions¶

PathHandler¶

PathHandler (motsart.common) is the central path utility used across all stages.

Configuration¶

All runtime entrypoints use Hydra-Zen-backed config stores. See Configuration.

Module summary¶

Module	Entry Point	Purpose
`complex_finder`	`python -m motsart.complex_finder.complex_finder`	Find reactant complexes
`path_guessers.rmsd_pp`	`python -m motsart.path_guessers.rmsd_pp.rmsd_pp_reaction_path_guesser`	Generate initial TS guesses
`path_guessers.ts_conf_sampler`	`python -m motsart.path_guessers.ts_conf_sampler`	Refine RMSD-PP guesses via RacerTS
`validator`	`python -m motsart.validator.base_validator`	Validate TS guesses + IRC
`learning.results_to_data_pkl`	`python -m motsart.learning.results_to_data_pkl`	Build AL training/eval data
`learning.rtsp_guesser`	`python -m motsart.learning.rtsp_guesser`	Multihead RTSP guess generation

Results directory structure¶

Each reaction R{rxn_id} has its own subtree:

results*/
└── R{rxn_id}/
    ├── r/
    │   ├── temp/
    │   ├── struct_xyzs/
    │   └── final_complexes/
    ├── p/
    ├── ts/
    │   ├── rmsd_pp/
    │   ├── racer_ts/
    │   └── learning/
    └── validation/
        ├── rmsd_pp/
        ├── racer_ts/
        └── learning/