Start with a scientific question, not a software chain

AI assistance in computational chemistry is most useful when it organizes a well-defined calculation: checking inputs, identifying missing decisions, preparing a handoff, and explaining outputs alongside their limitations. A fluent response is not a substitute for a structure, a kinetic mechanism, a thermodynamic database, or simulation data.

The six Skills in this guide address different stages and scientific questions. They should not be assembled into a default Cantera → pycalphad → Phonopy pipeline. Gas-phase ignition, finite-temperature alloy equilibrium, lattice vibrations, and reactive molecular-dynamics post-processing require distinct models and input contracts.

Choose a branch from the observable you need. For materials discovery, that might be a phonon spectrum or equilibrium phase fraction. For process optimization, it might be a mechanism-specific ignition-delay comparison. Record the question, physical assumptions, input provenance, and acceptance criteria before asking an agent to prepare work.

Related guides: Chemistry and Materials AI Skills · Chemistry AI Agent Stack.

Understand the Skill stack and its boundaries

These resources are instruction-based Skills, sometimes accompanied by helpers. They are not the upstream scientific packages, autonomous running agents, or hosted simulation APIs. Loading a Skill gives a host procedural guidance; it does not install a solver, supply calculation permissions, or guarantee execution. Host support also differs.

Resource Source-described role Required scientific input Output or handoff
Pymatgen Skill Structure validation, symmetry sensitivity, conversion planning, local hull analysis Structures or compatible computed-energy entries Reports, converted artifacts, query plans, manifests
ASE Skill Route atomistic workflow and calculator requests Task intent and minimum branch context Branch choice, missing inputs, delegation step
Phonopy Skill Organize finite displacements and phonon analysis Structure plus displaced-supercell forces or force constants Displacement layout, assembly status, phonon-band/DOS/thermal files
Cantera Skill Helper-guided homogeneous ideal-gas ignition Mechanism, composition, reactor constraint, conditions JSON report, four histories, mechanism snapshot
pycalphad Skill Helper-guided finite-temperature CALPHAD equilibrium Local TDB, components, phases, composition, conditions Equilibrium CSV and numerical-check report
ReacNetGenerator Skill Reactive-MD post-processing and output inspection Actual trajectories, atom mapping, cell context Species/reaction artifacts, reports, logs

Treat cross-resource handoffs below as proposed integrations, not tested interoperability. The upstream packages and any external force engines remain separate dependencies.

Prepare inputs and preserve atomic meaning

For a proposed structure-first workflow, use the Pymatgen Skill to inspect whether the input represents a periodic structure or non-periodic molecule. Establish units, coordinate mode, lattice, occupancy, disorder, and oxidation-state treatment. Preserve parser warnings rather than silently accepting corrections.

Conversion is a scientific decision, not merely a file-writing step. The Skill describes planning representation loss, writing to a new path, and checking relevant properties after round-tripping. Before a downstream handoff, compare species ordering, lattice, periodicity, coordinates, and site information. Its contact check has limits for very short lattice vectors, so a successful validation report is not exhaustive structural certification.

If the task needs symmetry, inspect sensitivity across explicit distance and angle tolerances. Do not tune tolerances merely to obtain a preferred space group.

The ASE top-level Skill then offers a routing contract: static, relaxation, MD, and NEB intentions go to ase/ase-workflows; calculator configuration goes to ase/ase-calculators. Mixed requests begin with the workflow branch. The router reports its choice, rationale, missing inputs, and next delegation. It does not execute calculations or prescribe backend parameters itself.

Branch one: phonons from structures and real forces

Choose Phonopy when the objective is lattice-vibration band structure, density of states, or thermal quantities. Required decisions include cell context, force provider, supercell, displacement amplitude, symmetry settings, and target outputs. Bands additionally need a defined path and sampling; DOS and thermal work need a mesh, with temperature settings for thermal analysis.

The proposed architecture is:

Validated structure → displacement tasks → separate force provider → checked force collection → force constants → requested analysis.

The Skill names providers such as VASP, Quantum ESPRESSO, and ML force fields, but that is not evidence of tested integration. Force execution and cluster submission belong elsewhere. During preparation, identify what forces are still needed; do not proceed to force-constant assembly without actual displaced-supercell forces or supplied force constants.

Keep an explicit displacement-to-force-file mapping. Missing files, changed atom order, or inconsistent units should block assembly. Report relevant long-range or non-analytic correction settings. Imaginary modes require investigation of convergence, supercell size, and setup before interpretation; neither their presence nor their absence alone settles material suitability.

Branch two: gas-phase ignition with Cantera

Choose the Cantera Skill for closed, adiabatic, homogeneous ideal-gas ignition under constant-volume or constant-pressure constraints. Its bundled helper is not a flame solver or general reactor-network builder.

Inputs include a kinetic mechanism, initial temperature and pressure, mole amounts, simulation horizon, output spacing, and solver controls. Custom mechanisms must be available as Cantera YAML. Confirm the ideal-gas phase and required species, preserve mechanism provenance, and choose the reactor constraint from the physical question rather than convenience.

The helper defines ignition delay as the time at the global maximum temperature derivative on a uniform output grid. It withholds the delay when heating is insufficient or the maximum is too near a time boundary. Four independent runs examine output spacing, solver controls, and horizon sensitivity.

Inspect the JSON report and all four CSV histories, including mass, elemental, and appropriate energy conservation checks. A null delay is an unresolved or nonigniting status under the documented criteria—not a number to estimate conversationally. Numerical resolution does not establish mechanism validity, and thermodynamic-range checks cover saved states rather than every internal integration state.

Branch three: CALPHAD equilibrium with pycalphad

Choose the pycalphad Skill for equilibrium phase fractions and compositions at one bulk elemental composition, one pressure, and an explicit finite-temperature list. Supply a local TDB with documented provenance and an appropriate assessed domain. The bundled ideal-cu-ni.tdb is hypothetical teaching data, not an assessed Cu-Ni database.

Specify components and candidate phases deliberately. The helper requires exactly N-1 independent elemental mole fractions and one dependent non-vacancy element; it does not silently normalize composition. Convert weight-based compositions before submission. Record phase exclusions because they can constrain the equilibrium result.

Outputs include report.json and phase-equilibria.csv. Phase fractions are molar. Multiple composition sets sharing a phase name are retained, which matters for miscibility gaps; do not collapse them into one composition.

Acceptance involves finite Gibbs energies, phase-fraction totals, reconstructed bulk composition, and sensitivity to increased phase-constitution sampling. These checks neither certify a global minimum nor validate the database experimentally. Inspect recorded solver composition adjustments before interpreting endpoint or ultratrace requests. Equilibrium results also do not predict precipitation rates or retained metastable microstructures.

Pymatgen's local computed-energy hull is a different analysis, conditional on compatible energies and competing entries—not a replacement for this finite-temperature TDB workflow.

Branch four: reactive-MD trajectory analysis

Choose ReacNetGenerator after a reactive-MD simulation has produced actual bond, dump, XYZ, or extended XYZ trajectories. The Skill does not generate those trajectories or validate their underlying force model.

Establish atom names and type mapping first. A LAMMPS data file may support inference, but ambiguous assignments require clarification. Check whether dump coordinates are scaled or Cartesian and whether cell information supports the intended periodic treatment. The separate reacnet-md-tools wrapper describes coordinate conversion for orthorhombic and triclinic cells.

The Skill routes routine LAMMPS dump work to rng-pipeline, lower-level requests to native reacnetgenerator, and existing species/reaction-output queries to rng-query. A local browser interface is a separate, explicitly requested option. Record the HMM choice rather than treating a quick initial pass as definitive analysis.

Inspect logs alongside generated species, reaction, and report artifacts. As a proposed scientific evaluation, compare important events with the underlying trajectory and examine sensitivity to analysis choices. A generated network should not automatically be relabeled as a validated gas-phase kinetic mechanism for Cantera.

Proposed example: a phonon-DOS pilot with stop gates

Suppose a researcher supplies a periodic input.cif and wants an exploratory phonon DOS. This is a proposed evaluation, not a completed calculation.

  1. Intake: preserve the original file and checksum. Request structure validation and a symmetry-sensitivity report through the Pymatgen Skill.
  2. Decision: resolve occupancy, coordinate, or parser ambiguities. Approve any conversion only after a representation-loss plan and round-trip checks.
  3. Preparation: ask the Phonopy Skill to organize displacement tasks using explicitly approved cell, supercell, amplitude, and mesh choices. Do not infer production convergence settings.
  4. Execution handoff: send displaced structures to a separately configured force provider. If ASE routing is useful, use its workflow/calculator split as a proposed preparation aid, not as an execution guarantee.
  5. Stop gate: missing forces, inconsistent ordering, or incompatible units block assembly. Request corrected artifacts rather than fill gaps.
  6. Analysis: assemble force constants and produce DOS only after completeness checks. Propose convergence comparisons and preserve unresolved scientific choices.

The expected package contains the input, validation records, displacement manifest, force artifacts, analysis settings, outputs, and limitations. No CALPHAD database, ignition mechanism, or reactive trajectory is needed for this branch.

Automate handoffs, not scientific judgment

A proposed agent workflow should maintain explicit states: inputs incomplete, preparation ready, execution authorized, outputs available, checks failed, and interpretation pending. Separate permission to prepare from permission to run calculations or access external services.

Pymatgen's Materials Project helper defaults to offline query planning; retrieval requires explicit approval, bounded criteria, and authentication. Materials Project is the external data service, while mp-api is its client—not a service supplied by the Skill.

On failure, retain diagnostics and failed checks. Refine unresolved ignition sampling, investigate equilibrium sensitivity, repair force collections, or clarify trajectory mappings according to the branch. Avoid unbounded retries and do not report failed checks as convergence.

Preserve software environments, settings, hashes, warnings, and parent–child artifact relationships. A Cantera snapshot hash identifies saved bytes and may differ between otherwise equivalent reruns; retain imported mechanism dependencies separately. Licence scope also matters: ASE's Skill declares MIT while its collection supplies LGPL text, leaving their relationship unresolved. Clarify that scope before redistribution rather than assuming repository visibility establishes rights.

Reader checklist and related guides

  • Is the scientific observable and correct branch explicit?
  • Are structures, forces, mechanisms, databases, and trajectories real supplied inputs?
  • Are units, composition basis, atom ordering, and periodicity recorded?
  • Are external dependencies and execution permissions separate from Skill loading?
  • Are failed checks and unresolved results preserved?
  • Are numerical consistency and scientific validity evaluated separately?
  • Can another researcher reconstruct the artifact lineage and assumptions?

For broader context, read agents and Skills and AI Skills for chemistry. Connect these branches to materials-science workflows or chemical-engineering workflows according to the question. Use the reproducible chemistry AI guide to plan records across preparation, execution, and interpretation.

Sources

Primary workflow contracts: ASE Skill instructions, Pymatgen Skill instructions, and Phonopy Skill instructions.

Branch definitions and output checks: Cantera Skill instructions, pycalphad Skill instructions, and ReacNetGenerator Skill instructions. The collection licence text supports the ASE licence-scope caution.