The workflow · Keep the chemical question in view

Start with the chemical region. An interpretable descriptor states its region, physical property and aggregation rule explicitly.

1. Choose the region

  • Element mode: sulfur in an additive series with a sulfur-based prior.

  • Fragment mode: a conserved catalytic core in an organocatalyst series.

  • Automated local mode: local environments without a predefined region.

Choose before inspecting test outcomes. Keep whole-molecule descriptors alongside local ones so that size, shape and exposure remain visible. See Choose a descriptor mode for the three modes.

2. Structure → wavefunction → surface → descriptors

Prepare structures from SMILES, optimize with xTB, generate the Gaussian wavefunction, and obtain ESP/ALIE/LEAE with Multiwfn. Extract whole-molecule and selected-region descriptors. Persistent sample names connect every stage.

surfacia smi2xyz -i sulfur_molecules.csv
# After quantum/Multiwfn stages produce your actual FullOption.csv:
surfacia extract-features -i FullOption.csv --mode 1 --element S

Download Case 2 · Sulfur-centered surfaces for six public sulfur structures. Record charge, multiplicity, conformer, method, solvent, density isovalue and units. Do not replace failed calculations with an unexplained zero.

The article's recovered Mode 1/2 production settings were GFN2-xTB with ALPB water, then a PBE1PBE/6-311G* single point with GD3 and SMD water, then Multiwfn at a 0.010 a.u. density isovalue. These are article settings, not a guarantee that CLI defaults reproduce the protocol. Review generated inputs before running.

3. Descriptors → model → explanation

With traceable targets and a fixed held-out set, fit a predefined feature block. Keep held-out outcomes out of feature selection and interpretation ranking. The configured end-to-end shortcut is:

surfacia workflow -i molecules.csv --mode 1 --element S --test-samples "6,12" --epoch 64 --cores 4

This requires installed external chemistry programs and known targets; it was not quantum-rerun for this manual. Inspect stages individually for a new system. --resume detects files; it does not validate scientific settings. workflow lists the implemented options.

4. Attribution → chemical evidence

Read the descriptor definition, the trend across molecules and original predictions together. Ask which molecular change would test the suggested relationship. See From model attribution to chemical meaning for a worked reading.

The public CLI provides an XGBoost/repeated-split workflow. The article's shared 128-split, dual-model OOF TreeSHAP/DeepSHAP, perturbation and blinded multi-LLM assessment are a separately frozen research analysis. workflow does not automatically run that assessment or reproduce the article benchmark.