chemiscope-python

Chemiscope in Python

Chemiscope is an interactive explorer for collections of molecules and materials. A dataset pairs structures (atomic geometries) with properties (scalars/vectors per structure or per atom). Three linked panels stay in sync: a 2D property map (Plotly scatter), a 3D structure viewer (3Dmol), and an info panel.

Requires pip install chemiscope. Reading structures needs a backend such as ase (pip install ase), or chemfiles / MDAnalysis / stk.

Reference files (load on demand)

• settings-reference.md — full settings dict schema (every map / structure field, defaults, accepted values, palettes). Read it when hand-authoring settings beyond quick_settings.
• shapes-reference.md — full shapes spec (all kinds, parameters, helper signatures). Read it when building custom shapes beyond the cookbook recipes below.

Choosing the right output

Pick the delivery mechanism before writing code:

There is no one-call Python API that emits a self-contained HTML; the standalone route is "JSON + prebuilt chemiscope_standalone.html" (concatenation, uncompressed JSON only). See "Other delivery mechanisms" at the bottom for Sphinx/Streamlit/headless detail.

show / show_input / streamlit.viewer / the .. chemiscope:: directive all accept mode= "default" (map + structure) | "structure" (3D only) | "map" (map only).

Minimal recipe

import ase.io
import chemiscope

structures = ase.io.read("trajectory.xyz", ":")  # ":" reads ALL frames

chemiscope.show(
    structures=structures,
    properties=chemiscope.extract_properties(structures),  # pull props from the files
    metadata={"name": "My dataset"},
    settings=chemiscope.quick_settings(x="energy", y="volume", map_color="energy"),
)

Swap show(...) for write_input("out.json.gz", ...) (same args) to save instead. create_input(...) returns the dataset dict without writing — useful for inspection.

Input structures

structures is a list of any supported object (don't mix types):

ASE emits a cell only for periodic systems (non-zero lattice).

Hand-built structure dict

{
    "size": 3,                       # required: number of atoms
    "names": ["O", "H", "H"],        # required: per-atom symbol/name (len == size)
    "x": [0.0, 0.76, -0.76],         # required: Cartesian coords (len == size)
    "y": [0.0, 0.59, 0.59],          # required
    "z": [0.0, 0.0, 0.0],            # required
    "cell": [a1,a2,a3, b1,b2,b3, c1,c2,c3],  # optional: flattened 3x3, row-major
    # optional: "bonds": [[i,j,order],...], "elements", "chains", "resnames",
    #           "resids", "hetatom"
}

Large datasets: external structures

external = chemiscope.write_external_structures(structures, prefix="structure")
chemiscope.write_input("dataset.json.gz", structures=external, properties=...)
# Structure files must be shipped alongside the JSON; such a dataset can't be opened
# on chemiscope.org (which needs everything inline).

Properties

A dict keyed by name. Full form (preferred for sharing — units/description matter) or shorthand (target auto-deduced by length):

properties = {
    "energy": {                      # full form
        "target": "structure",       # "structure" (one per frame) or "atom" (one per atom)
        "values": [1.2, 0.9, 1.5],   # list, 1D np.ndarray, or 2D np.ndarray
        "units": "eV", "description": "DFT energy",
    },
    "volume": [12.0, 12.1, 11.9],    # shorthand
}

• A 2D numpy array (N×M, M>1) becomes M properties name[1] … name[M] (PCA components, eigenvalues, tensor entries…). Reference them as x="pca[1]".
• target="atom" values length = total atoms across all structures (or # selected environments). NaN hides that point on the map.

Extract from the structures

props = chemiscope.extract_properties(structures)                  # everything
props = chemiscope.extract_properties(structures, only=["energy", "forces"])
props = chemiscope.extract_properties(structures, environments=envs)  # atom props for a subset

For ASE: structure props come from atoms.info, atom props from atoms.arrays; attached-calculator energy/forces are injected automatically (so forces show up as an atom property named "forces"). Only props present in every frame are kept.

Parameterized (multidimensional) properties

properties = {"spectrum": {"target": "structure", "values": spectra_2d,
                           "parameters": ["freq"]}}
parameters = {"freq": {"values": freq_grid_1d, "units": "cm^-1", "name": "frequency"}}
chemiscope.write_input("out.json.gz", structures=s, properties=properties,
                       parameters=parameters)

Environments (atom-centered properties)

To color/select per atom, declare which atoms carry data and their display cutoff, as (structure_id, center_id, cutoff) tuples:

environments = chemiscope.all_atomic_environments(structures, cutoff=3.5)  # all atoms
environments = [(0, 4, 3.5), (0, 7, 3.5), (2, 0, 4.0)]                     # manual subset

Pass environments=... and use target="atom" in the settings for per-atom views.

Settings (common case)

settings = chemiscope.quick_settings(
    x="energy", y="volume", z=None,        # map axes (property names; use name[i] for components)
    map_color="energy", size=None, symbol=None,
    structure_color=None,                  # color atoms by a property
    target="structure",                    # "atom" for per-environment view
    trajectory=False,                      # True: join map points + keep 3D orientation
    periodic=False,                        # True: show unit cell + 3x3x3 supercell
)

For full control (every map/structure field, palettes, etc.), read settings-reference.md and pass a settings dict directly.

Shapes (common case)

Pass shapes={name: definition}, then activate via the structure setting "shape": "name". Helpers build definitions for you:

chemiscope.ase_vectors_to_arrows(structures, "forces", scale=1, radius=0.15)
chemiscope.ase_tensors_to_ellipsoids(structures, "alpha", force_positive=True, scale=0.2)

For custom meshes, batched spheres/cylinders, combined shapes, layered per-atom parameters, and the full kind list, read shapes-reference.md.

Cookbook

Forces / vectors as arrows

chemiscope.write_input(
    "trajectory.json.gz",
    structures=structures,
    properties=chemiscope.extract_properties(structures),
    shapes={"forces": chemiscope.ase_vectors_to_arrows(structures, "forces", scale=1, radius=0.15)},
    settings={
        "structure": [{"keepOrientation": True, "playbackDelay": 100, "shape": "forces"}],
        "map": {"joinPoints": True},
    },
)

Works for any N×3 atom array (velocities, dipoles, displacements…).

Atom-centered property: color atoms + labels

chemiscope.show(
    structures=structures,
    properties={"charge": {"target": "atom", "values": charges_flat, "units": "e"}},
    environments=chemiscope.all_atomic_environments(structures, cutoff=3.5),
    settings={
        "target": "atom",
        "structure": [{"color": {"property": "charge", "palette": "bwr"},
                       "atomLabels": True, "labelsProperty": "charge"}],
    },
)

Tensors as ellipsoids (e.g. polarizability)

chemiscope.write_input(
    "tensors.json.gz", structures=structures,
    properties=chemiscope.extract_properties(structures, only=["alpha"]),
    shapes={"alpha": chemiscope.ase_tensors_to_ellipsoids(structures, "alpha",
                                                          force_positive=True, scale=0.2)},
    settings={"structure": [{"atoms": False, "shape": "alpha"}]},
    environments=chemiscope.all_atomic_environments(structures),
)

Custom mesh shape, per-atom scaling

cube = {
    "kind": "custom",
    "parameters": {
        "global": {"vertices": [[-1,-1,-1],[1,-1,-1],[-1,1,-1],[1,1,-1],
                                 [-1,-1,1],[1,-1,1],[-1,1,1],[1,1,1]]},  # hull auto-computed
        "atom": [{"scale": 0.3, "color": 0x44AAFF} for _ in range(total_atoms)],
    },
}
chemiscope.write_input("mesh.json.gz", structures=structures, shapes={"cube": cube},
                       settings={"structure": [{"shape": "cube"}]})

Trajectory / periodic / biomolecules

settings = chemiscope.quick_settings(x="time", y="energy", trajectory=True)  # path + fixed view
settings = chemiscope.quick_settings(periodic=True, target="atom")           # cell + supercell

import MDAnalysis as mda
u = mda.Universe("protein.pdb")
chemiscope.show(structures=u.atoms, mode="structure",
                settings=chemiscope.quick_settings(structure_settings={"cartoon": True}))

Metadata

metadata = {"name": "Dataset title",          # required (bare minimum)
            "description": "...", "authors": ["Ada Lovelace"],
            "references": ["doi:10.xxxx/yyyy"]}

Automatic dimensionality reduction

chemiscope.explore(structures, featurizer="pet-mad-1.0",
                   properties=chemiscope.extract_properties(structures),
                   environments=chemiscope.all_atomic_environments(structures))
# write straight to a file: explore(..., write_input="out.json.gz")
# custom model: featurizer = chemiscope.metatomic_featurizer(model="model.pt")

Other delivery mechanisms

Self-contained offline HTML

chemiscope.write_input("data.json", structures=s, properties=p)  # uncompressed JSON!

Then download chemiscope_standalone.html from https://chemiscope.org/chemiscope_standalone.html and concatenate:

cat chemiscope_standalone.html data.json > out.html   # opens in any browser, no server

Only uncompressed .json (not .json.gz) works with the standalone HTML.

Sphinx documentation

Add the extension in conf.py: extensions = ["chemiscope.sphinx"]. Then in an .rst:

.. chemiscope:: ../datasets/showcase.json.gz
    :mode: structure       # default | structure | map (optional, default "default")
    :warning_timeout: -1   # ms; 0 = persistent, -1 = disable (optional)

The path is to a .json/.json.gz file relative to the .rst. The widget loads data dynamically, so serve the built docs over HTTP (e.g. python3 -m http.server in build/html); opening the raw file won't work. (For sphinx-gallery examples, assigning a show() widget to a variable named ___ auto-generates the directive via the bundled ChemiscopeScraper.)

Streamlit app (pip install chemiscope[streamlit])

import chemiscope, ase.io, streamlit as st
structures = ase.io.read("structures.xyz", ":")
dataset = chemiscope.create_input(structures)
chemiscope.streamlit.viewer(dataset, mode="structure")   # run with: streamlit run app.py

viewer(dataset, *, settings=None, mode="default", no_info_panel=False, width="stretch", height=550, key="chemiscope_viewer", selected_index=None, on_select=None, on_settings_change=None). on_select receives the selected structure index (or None); selected_index (e.g. bound to st.session_state) drives the selection externally.

Headless / scripted images

w = chemiscope.show(structures=structures, properties=props, settings=settings)
await w.save_structure_image("frame.png")   # await the futures in a notebook
await w.save_map_image("map.png")
await w.save_structure_sequence([0, 1, 2], ["a.png", "b.png", "c.png"])
# fully headless (no notebook), Playwright-based:  from chemiscope import headless

Gotchas & validation cheat-sheet

• settings must be a JSON-serializable dict (round-tripped through JSON).
• target="atom" / atom-coloring requires environments.
• mode="default"/"map" needs ≥2 properties for the map.
• Properties-only dataset (no structures) requires every property target="structure".
• Atom-property length = total atoms; structure-property length = number of structures.
• Change a widget's view by reassigning .settings with a complete dict (in-place nested mutation is ignored).
• frames= / meta= are deprecated aliases for structures= / metadata=.
• extract_properties drops properties not present in every structure (with a warning).

Using this skill elsewhere

This skill is self-contained, portable, and harness-agnostic Markdown. In the chemiscope repository the canonical copy lives at docs/skills/ (this SKILL.md plus the two reference files); Claude Code also picks it up via the symlink .claude/skills/chemiscope-python -> ../../docs/skills.

To reuse it in another project or agent:

• Claude Code — copy (or symlink) the three files into a named folder under ~/.claude/skills/<name>/ (available in every project) or <project>/.claude/skills/<name>/ (scoped to one project).
• Other agents — point the agent at these files, e.g. from an AGENTS.md pointer or by referencing the path directly. The YAML frontmatter is Claude-specific metadata and is harmless to other tools.

Either way, pip install chemiscope (plus a backend like ase) in the working environment.