PhysicsNeMo Discover
Find suitable PhysicsNeMo models, datapipes, and examples for physics-ML tasks.
The skill explicitly limits itself to read-only discovery, forbids code writing, and requires live path discovery and verification; no malware, credential theft, or covert exfiltration is evident. However, it lacks user confirmation, rollback guidance, and complete data-flow disclosure, and it may perform an external Git clone without documenting network consent, integrity verification, or stronger temporary-directory isolation, so points are deducted.
SKILL.md defines repository resolution, search, read, ls verification, fail-and-drop behavior, and out-of-scope abstention, making the baseline workflow coherent. There is no committed executable test suite or key-path reproduction evidence, external cloning can fail, and abnormal-input feedback is only partly specified; the score is therefore capped by static calibration and reduced.
The audience, SciML/AI4Science scenarios, non-fit ranges, trigger conditions, data-shape follow-ups, and a concrete Chinese-language question are clear. However, Chinese output support is not declared, and the headless fallback depends on GitHub when no local clone exists, creating possible mainland-China reachability limitations, so points are deducted.
The documentation is layered across core instructions, output and abstention contracts, TAXONOMY, and RECIPES; dependencies are limited, and author/license metadata is present. Nevertheless, the skill lacks an explicit version, changelog, clear maintenance owner, and update path in the supplied skill files, while BENCHMARK reports missing recommended Instructions/Examples sections, so points are deducted.
The skill turns repository navigation into a verifiable menu of models, datapipes, examples, and documentation, and directly addresses the discovery task by prohibiting memory-based citations. The supplied BENCHMARK reports positive and negative tasks plus effectiveness metrics, but it does not provide independently reproducible representative outputs; the static ceiling therefore applies and points are deducted.
RECIPES and the evaluation files provide auditable search procedures, task definitions, and reported results, and the assessed revision is pinned. However, no committed test suite, CI coverage, or raw execution logs are supplied, so the evaluation conclusions remain primarily author-reported and cannot support a high verifiability score.
- Before first use, confirm that external GitHub access is permitted and review the temporary-directory, proxy, and credential environment.
- Do not treat the BENCHMARK PASS, accuracy, or security results as independently reproduced evidence in this static review.
- In mainland-China network-restricted environments, prefer a local PhysicsNeMo clone because live discovery may otherwise fail.
What does this skill do, and when should you use it?
PhysicsNeMo Discover is NVIDIA-authored guidance for navigating PhysicsNeMo in SciML and AI4Science workflows. It uses the current PhysicsNeMo repository to discover models, datapipes, examples, and documentation based on the user’s task and data shape. The skill treats model, datapipe, training strategy, and configuration as composable axes rather than presenting one example as the prescribed solution. It provides read-only navigation guidance and does not write training code or handle installation and environment setup.
Reads current PhysicsNeMo exports, module documentation, example READMEs, documentation, and configuration files; uses live repository search, Glob, Read, Grep, and Bash path verification to locate model families, datapipes, examples, and supporting documents; shallow-clones NVIDIA/physicsnemo into a temporary directory when no local clone is available, for read-only path discovery; and produces candidate menus, reference examples, reading order, and scope or abstention guidance.
- A SciML researcher needs to choose PhysicsNeMo models for regular-grid, spherical-grid, or unstructured-mesh data.
- An AI4Science engineer wants to compare model families and datapipes independently instead of copying a single example.
- A user needs to verify whether a class, file, or example still exists in the current PhysicsNeMo repository.
- A team wants an agent to route surrogate, forecasting, downscaling, physics-informed, inverse, or generative physical-system tasks to relevant repository resources.
What are this skill's strengths and limitations?
- Official NVIDIA-authored guidance focused on PhysicsNeMo SciML and AI4Science discovery.
- Requires live discovery and path verification each turn, reducing stale-path and rename errors.
- Separates model, datapipe, training-strategy, and configuration choices for comparison.
- Clearly excludes code authoring, installation setup, and general non-physics ML tasks.
- Requires shell, filesystem, and network access for live discovery or shallow cloning.
- Does not write training code or provide installation and environment setup guidance.
- The supplied source does not provide a skill-specific test suite or concrete benchmark results.
- Results depend on access to the current PhysicsNeMo repository and successful path verification.
How do you install this skill?
Install the skill from the NVIDIA/skills catalog with the skills CLI:
npx skills add nvidia/skills --skill physicsnemo-discover --yes
The README does not document the final local installation directory; the CLI handles the destination.
How do you use this skill?
After installation, give the agent a concrete PhysicsNeMo discovery request, for example: "Based on my physical-field data shape and forecasting task, find suitable model families, datapipes, and reference examples in the current PhysicsNeMo repository, with absolute paths." Include the task and data shape so the skill can perform live discovery and return read-only navigation guidance.