Chem Ts Optimization

TS Optimization with Sella

Optimize a transition-state guess and check whether it is a first-order saddle point.

Scope

• Domain: molecular chemistry only (non-periodic systems).
• Trigger: user has a TS guess and needs TS optimization plus frequency validation.
• Exclusions: periodic diffusion/path workflows (use chem-neb-barrier instead).

Tool

optimize_ts_sella.py

Runs Sella TS optimization followed by finite-difference vibrations.

Use with MACE

# Env: mace-agent
python .agents/skills/chem-ts-optimization/scripts/optimize_ts_sella.py \
  --ts_guess ts_guess.xyz \
  --model_type mace \
  --model_name MACE-OFF23-small \
  --fmax 0.02 \
  --steps 500 \
  --imag_cutoff_cm1 -50.0 \
  --output_dir results/ts_opt

Use with FAIRChem (UMA)

# Env: fairchem-agent
python .agents/skills/chem-ts-optimization/scripts/optimize_ts_sella.py \
  --ts_guess ts_guess.xyz \
  --model_type fairchem \
  --model_name uma-s-1p1 \
  --task_name omol \
  --fmax 0.02 \
  --steps 500 \
  --imag_cutoff_cm1 -50.0 \
  --output_dir results/ts_opt

Arguments

• --ts_guess: required TS guess geometry (XYZ supported by ASE I/O).
• --model_type: required backend (mace or fairchem).
• --model_name: optional model identifier/checkpoint.
• --task_name: optional model head/task (for UMA molecular runs use omol).
• --device: auto|cpu|cuda (default auto).
• --fmax: Sella convergence threshold in eV/A (default 0.02).
• --steps: maximum TS optimization steps (default 500).
• --vib_delta: finite-difference displacement in A (default 0.01).
• --vib_nfree: finite-difference stencil size (2 or 4, default 2).
• --imag_cutoff_cm1: imaginary mode cutoff in cm^-1 (default -50.0).
• --keep_vib_cache: optional flag to keep vibration cache files in output_dir/vib.
• --output_dir: required output directory.

Outputs

• ts_optimized.xyz: optimized TS geometry.
• ts_opt.traj: TS optimization trajectory.
• ts_opt.log: optimizer log.
• ts_optimization_results.json: run summary and pass/fail decision.
• vib/ cache files only when --keep_vib_cache is set.

ts_optimization_results.json fields include:

• run/model metadata
• convergence (sella_converged, optimization_steps, max_force_eV_per_A)
• vibrational data (all_frequencies_cm1, imaginary_modes)
• classification (n_imag_below_cutoff, is_first_order_saddle)

TS Pass Criterion

A structure is accepted as first-order saddle only if:

• exactly one frequency satisfies frequency < imag_cutoff_cm1

Default criterion: exactly one mode below -50 cm^-1.

Model Guidance

• Recommended for molecules:
• MACE-OFF23-small / MACE-OFF23-medium
• uma-s-1p1 with --task_name omol
• Use the same backend/model/head across reactant/product/TS optimization and TS validation.

Prerequisites And Constraints

• Activate mace-agent or fairchem-agent depending on backend.
• Script enforces pbc=False (non-periodic only).
• TS guess quality matters; poor guesses can converge to minima or higher-order saddles.

Examples

See examples/ directory for sample inputs and outputs. ---

Author: Juno Nam Contact: GitHub @recisic

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: learningmatter-mit
• Source: learningmatter-mit/AtomisticSkills
• License: MIT
• Homepage: https://arxiv.org/abs/2605.24002

Install and usage instructions live in the source repository linked above.