Skill: catalyst SAR screening

SAC structure–activity screening with a fixed pipeline: build graphene M–N–C POSCARs from the embedded catalog, evaluate metrics with FAIRChem UMA only, analyze SAR trends, and write a lean visual report.

All skill files live flat under this directory (no data/ or examples/ subfolders).

HARD LOCK

• Energy engine: Catalyst-Design-Agent FAIRChem UMA uma-s-1p1, task oc20
• Runtime: conda env catagent
• Forbidden: tabular / heuristic / lookup / other MLIPs; skipping UMA when HF_TOKEN or Hugging Face is missing; returning committed demo shells (metal_center_dissolution_*) as the user answer

When blocked: ask the user for HF_TOKEN and/or HF_ENDPOINT (e.g. https://hf-mirror.com). Do not substitute another method.

Example user prompt

Use this as the canonical prompt shape (replace the token placeholder; never commit a real token):

请使用 catalyst_sar_screening 这个 skill，基于 UMA 模型评估载体为石墨烯的
M–N4 的电化学稳定性；电化学稳定性用溶解电位量化；金属中心搜索空间包括
Mn、Fe、Cu；分析金属中心与电化学稳定性的关系并生成可视化报告。
huggingface 的 API 使用 <HF_TOKEN_PLACEHOLDER>；
运行代码前 export HF_ENDPOINT=https://hf-mirror.com。

Mapped run:

import os

os.environ["HF_TOKEN"] = "<HF_TOKEN_PLACEHOLDER>"  # from the user message only
os.environ["HF_ENDPOINT"] = "https://hf-mirror.com"

from catalyst_sar_screening.kernel import check_uma_readiness, run_pipeline

ready = check_uma_readiness()
if not ready["ok"]:
    raise SystemExit(ready["ask_user"])

result = run_pipeline(
    ["Mn-N4", "Fe-N4", "Cu-N4"],
    workdir="outputs/sac_mn_fe_cu_udiss",
    metrics=["dissolution"],
    min_dissolution=0.0,
)
# Present ONLY these paths to the user (never demo shells in this skill dir).
deliverables = result["deliverables"]

Fixed pipeline

1. Map the user request → descriptions + metrics (example above → ["Mn-N4","Fe-N4","Cu-N4"], metrics=["dissolution"]).
2. Set HF_TOKEN / HF_ENDPOINT from the user if provided (do not hardcode).
3. check_uma_readiness(); if not ok → stop and ask.
4. run_pipeline(...) builds POSCARs from contcar_catalog.json (embedded POSCAR text; exact lookup first, else derive). Do not read an external chem/ tree.
5. Evaluate with UMA only; analyze SAR; write report/figures under workdir.
6. Present only result["deliverables"].

Module layout

skills/catalyst_sar_screening/
├── SKILL.md
├── kernel.py
├── contcar_catalog.json              # minimal graphene M–N4 slab POSCAR texts
├── build_example.py                  # regenerate demo shells
├── metal_center_dissolution_*.json   # synthetic demos — NOT user outputs
├── metal_center_dissolution_*.md
└── metal_center_dissolution_*.html

The catalog is intentionally limited to graphene / pyridineN / slab entries for the public skill fixture. It is not an experimental dataset release.

Import

from catalyst_sar_screening.kernel import check_uma_readiness, run_pipeline

Backend setup

Pinned runtime packages for conda env catagent:

• fairchem-core==2.19.0
• pandas==3.0.2
• pymatgen==2026.3.23

conda create -n catagent python=3.11 -y
conda activate catagent
python -m pip install ase numpy matplotlib pandas==3.0.2 pymatgen==2026.3.23
python -m pip install fairchem-core==2.19.0 fairchem-data-oc==1.0.2 torch
export HF_TOKEN=...        # ask user if unset — never commit real tokens
export HF_ENDPOINT=https://hf-mirror.com

Deliverables

Only paths under the run workdir listed in result["deliverables"]:

<workdir>/
├── catalyst_sar_report.md
├── catalyst_sar_dashboard.html
├── summary.json
└── figures/
    ├── fig01_*.png
    └── structures_collage.png

Do not return committed demo shells such as:

skills/catalyst_sar_screening/metal_center_dissolution_*

Developer demos

Synthetic placeholder HTML/Markdown/JSON live flat in this skill directory for maintainers. They must not include unpublished numeric screening results or figure PNGs. Regenerate text shells with:

uv run python skills/catalyst_sar_screening/build_example.py

Analyst checklist

• ran run_pipeline for this request's metals/metrics
• every presented file is under that run's deliverables
• no committed metal_center_dissolution_* demo was attached
• report includes model + statistical figures + structure collage + SAR insights
• no real HF token was written into files or commits