input-skill

name

input-skill

description

Generate Gaussian and ORCA quantum-chemistry input files from user-provided SMILES strings using RDKit for single-conformer 3D structure generation, Multiwfn for geometry injection, resource-aware templates, and a curated benchmark rule library. Use when the user asks for Gaussian .gjf/.com or ORCA .inp input generation, SMILES-to-input workflows, automatic method/basis recommendation, or computational task setup with cores, memory, charge, multiplicity, and task type.

Input Skill

Use this skill to turn a SMILES string into selected Gaussian and/or ORCA input files with an auditable method/basis recommendation.

Required Inputs

Always collect these before running the pipeline:

• SMILES string.
• Output program: gaussian, orca, or both.
• Task: sp, opt, freq, opt-freq, tddft, nmr, interaction-sp, or high-accuracy-sp.
• Number of CPU cores.
• Total memory, for example 8GB or 64000MB.
• Net charge.
• Spin multiplicity.

Do not silently infer charge or multiplicity from SMILES. You may mention the formal charge after generation if RDKit reports a mismatch.

When the user has not chosen a program, ask whether they want Gaussian only, ORCA only, or both. Recommend both when they are comparing engines or preparing a portable starting point.

Quick Start

If RDKit is missing, ask the user before running setup:

python3 skills/input-skill/scripts/input_pipeline.py --setup

Ask interactively for missing inputs:

python3 skills/input-skill/scripts/input_pipeline.py --interactive

Generate both Gaussian and ORCA inputs:

python3 skills/input-skill/scripts/input_pipeline.py \
  --smiles 'c1ccccc1' \
  --task opt-freq \
  --cores 4 \
  --memory 8GB \
  --charge 0 \
  --multiplicity 1 \
  --program both

Generate only one program by changing --program to gaussian or orca.

Use --name <basename> to control output filenames. Use --workdir <dir> to choose where the timestamped input-skill-run/ directory is created.

Workflow

1. Read references/benchmark-library.yml when you need to explain or audit method/basis selection.
2. Run scripts/input_pipeline.py --setup only after the user approves installing RDKit into the skill-local .venv.
3. Ask the user for any missing required inputs, especially the output program choice.
4. Run the pipeline with all required inputs. It generates a single RDKit ETKDGv3 conformer and optimizes it with MMFF when available, otherwise UFF.
5. Prefer normal Multiwfn generation. The pipeline writes structure.xyz and only the selected program templates, then drives Multiwfn gi and/or oi -> -100 template injection.
6. Return the generated .gjf and/or .inp, recommendation.md, and manifest.json paths. Mention any warnings from manifest.json.

Resource Semantics

Interpret user memory as total job memory.

• Gaussian receives %nprocshared=<cores> and %mem=<total memory>.
• ORCA receives %pal nprocs <cores> end and %maxcore floor(total_MB * 0.75 / cores).

If per-core ORCA memory is below 256 MB, report the warning and ask the user to increase total memory or reduce cores before serious production runs.

Output Contract

The pipeline writes a timestamped output directory containing:

• source.smiles
• structure.xyz
• template.gjf and <name>.gjf when Gaussian is selected.
• template.orca.inp and <name>.inp when ORCA is selected.
• recommendation.md
• manifest.json
• multiwfn_gaussian.log and/or multiwfn_orca.log

The manifest records inputs, RDKit version, random seed, charge/multiplicity, resources, selected benchmark rule, citations, Multiwfn path, commands, outputs, and warnings.

Scientific Guardrails

• Treat generated 3D geometry as a starting structure, not a validated conformational search.
• Warn on unassigned stereocenters, formal-charge mismatch, transition metals, low memory, and Multiwfn fallback generation.
• For transition-metal or unusual spin-state systems, do not present the rule-library choice as generally benchmarked.
• For high-accuracy-sp, Gaussian output is a conservative DFT fallback when the selected ORCA method has no Gaussian equivalent.