Materials Candidate Screen

Skill workflow-skill many-tools-to-one-skill Proprietary (Paramus) — guidance only; execution performed by Paramus tools under their own license gating.
provenanceai generated statuspartial
Screen a materials question across public structure databases - the OPTIMADE federation and Materials Project - and return a candidate table with properties and per-row provenance.When to use: a user asks “which known materials have composition/property X” or wants an evidence-backed shortlist before any simulation.Multi-tool WORKFLOW with cross-database retrieval and provenance judgment.
Procedure7 steps · 9 orchestrated tools
1
Discover tools via search/get_schema ("optimade search", "materials project
2
Scope the query: translate the question into element/composition filters and a property
3
Query the federation with cloud_optimade_search_structures; fetch full records for
Toolscloud_optimade_search_structures
4
Query Materials Project with cloud_mp_search_materials, then pull properties per hit
Toolscloud_mp_search_materials
5
Enrich shortlisted candidates as the question demands: cloud_mp_get_bandstructure for
Toolscloud_mp_get_bandstructure
6
Reconcile: merge rows that refer to the same structure across sources (match on
7
Rank against the user's criterion and emit the table: material, key properties, source
You get
deduplicated candidate table (material, properties, source + entry id per row),

Overview

Most materials questions are answered twice if you let them be: once by Materials Project, once by the OPTIMADE federation — with overlapping but non-identical coverage and formats. This skill turns a composition/property question into one deduplicated candidate table in which every row carries its source database, entry id and the properties actually retrieved. The value is the retrieval-and-reconciliation logic, not any single query.

Guidance vs execution: all database access goes through the tested Paramus cloud connectors. Do not scrape provider APIs by hand or merge entries without checking structure identity.

When to Use

  • "Which known materials match this composition space / property window?" — a retrieval question over public data, answered with provenance
  • Building a candidate shortlist (with literature/DFT-computed properties already in the databases) before deciding what merits further work

Do not use for: QM screening of chosen compositions (new band-structure or DFT calculations on a picked candidate set — that is paramus-bandgap-semiconductor-screen), or predicting properties of hypothetical compounds not present in any database. This skill's trigger is cross-database retrieval and ranking with provenance; the trigger there is computation on compositions the user has already chosen.

Domain Judgment

  • Provenance is part of the answer: a property without its source entry id is not a screening result. Materials Project values are computed (mostly DFT-PBE); OPTIMADE rows vary by provider — say which kind of evidence each row is.
  • Deduplicate on structure, not formula: polymorphs are different candidates; the same polymorph in two databases is one candidate with two provenance entries.
  • Absence differs across databases: "not in MP" is a coverage statement, not a stability statement — check the phase diagram before calling a composition unstable.
  • Ranking criteria must be the user's: state the window/weights explicitly and flag rows whose deciding property is missing rather than interpolating.

Failure modes

  • A provider or the MP connector is unreachable → deliver the partial table with the missing source marked; never fill the gap from memory.
  • Zero hits → loosen exactly one filter at a time and report which relaxation produced hits; a bare "nothing found" is not an acceptable output.
  • Conflicting property values across sources for the same structure → show both values with their provenance; do not average.
  • Structure reconciliation ambiguous (near-identical lattices) → keep both rows and flag the ambiguity rather than merging silently.

Example

Asked: "Known binary zinc chalcogenides with a computed band gap in the 1–4 eV window — give me a table." The skill lists OPTIMADE providers, queries the federation for Zn–O/S/Se binaries, and queries Materials Project in parallel. The reconciled table includes wurtzite ZnO as mp-2133 with its MP-computed band gap of about 0.7 eV — which falls outside the window and is shown flagged as a known PBE underestimate (experimentally ~3.3 eV), while ZnS and ZnSe rows pass inside the window. Every row carries its source (OPTIMADE provider or Materials Project) and entry id; the report notes that JARVIS-DFT and NOMAD coverage is in implementation (jarvis-dft, nomad) and would widen the candidate set.

Input

composition space (elements/formula pattern), property window or ranking

Output

deduplicated candidate table (material, properties, source + entry id per row),