Advanced dataset exploration

The previous example introduced chemiscope.explore() and how to use it for automatic exploration of dataset. In this example, we’ll show some more representations and give additional featurizers you can use in your own code.

import os

import ase.io
import numpy as np
import requests

import chemiscope


def fetch_dataset(filename, base_url="https://zenodo.org/records/12748925/files/"):
    """Helper function to load the pre-computed examples"""
    local_path = "data/" + filename
    if not os.path.isfile(local_path):
        response = requests.get(base_url + filename)
        with open(local_path, "wb") as file:
            file.write(response.content)

Example with MACE-OFF and t-SNE

In this part, we are going to define another featurize function that runs calculation of desciptors with MACE-OFF and uses t-SNE for the dimensionality reduction.

The dependencies for this example can be installed with the following command:

pip install mace-torch scikit-learn

Let’s import the necessary libraries.

from dscribe.descriptors import SOAP  # noqa
from mace.calculators import mace_off  # noqa
from sklearn.manifold import TSNE  # noqa

/home/runner/work/chemiscope/chemiscope/.tox/docs/lib/python3.11/site-packages/e3nn/o3/_wigner.py:10: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  _Jd, _W3j_flat, _W3j_indices = torch.load(os.path.join(os.path.dirname(__file__), 'constants.pt'))

Load the dataset, in our example we are reading the organic molecules.

qm9_frames = ase.io.read("data/explore_qm9.xyz", ":")

Now, we are defining a featurize function. As on the previous example, it should return the reduced data.

def mace_off_tsne(frames, environments):
    if environments is not None:
        raise ValueError("'environments' are not supported by this featurizer")

    # At first, we initialize a mace_off calculator:
    descriptor_opt = {"model": "small", "device": "cpu", "default_dtype": "float64"}
    calculator = mace_off(**descriptor_opt)

    # Calculate MACE features for each frame
    descriptors = []
    for frame in frames:
        structure_avg = np.mean(
            # Only use invariant descriptors (no rotational components)
            (calculator.get_descriptors(frame, invariants_only=True)),
            axis=0,  # Average the descriptors over all atoms in the frame
        )
        descriptors.append(structure_avg)
    descriptors = np.array(descriptors)

    # Get number of jobs for parallelisation
    n_jobs = min(len(frames), os.cpu_count())

    # Apply t-SNE
    perplexity = min(30, descriptors.shape[0] - 1)
    reducer = TSNE(n_components=2, perplexity=perplexity, n_jobs=n_jobs)
    return reducer.fit_transform(descriptors)

We can also extract the additional properties, for example, dipole moment.

properties = chemiscope.extract_properties(qm9_frames, only=["mu"])

Provide the created featurizer and the properties to chemiscope.explore().

cs = chemiscope.explore(qm9_frames, featurize=mace_off_tsne, properties=properties)

Downloading MACE model from 'https://github.com/ACEsuit/mace-off/blob/main/mace_off23/MACE-OFF23_small.model?raw=true'
The model is distributed under the Academic Software License (ASL) license, see https://github.com/gabor1/ASL
 To use the model you accept the terms of the license.
ASL is based on the Gnu Public License, but does not permit commercial use
Cached MACE model to /home/runner/.cache/mace/MACE-OFF23_small.model
Using MACE-OFF23 MODEL for MACECalculator with /home/runner/.cache/mace/MACE-OFF23_small.model
Using float64 for MACECalculator, which is slower but more accurate. Recommended for geometry optimization.
/home/runner/work/chemiscope/chemiscope/.tox/docs/lib/python3.11/site-packages/mace/calculators/mace.py:128: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  torch.load(f=model_path, map_location=device)

Here we display the visualisation of the pre-computed data using the described function for 6k structures taken from the QM9 dataset. The map is zoomed in to highlight a cluster of zwitterions grouped together by application of the previously defined function with MACE-OFF and t-SNE.

fetch_dataset("mace-off-tsne-qm9.json.gz")
chemiscope.show_input("data/mace-off-tsne-qm9.json.gz")

×

Example with MACE-MP0 and t-SNE

We will define another featurize function that uses MACE-MP0 to calculate the descriptors and t-SNE for the dimensionality reduction.

Firstly, import mace library.

from mace.calculators import mace_mp  # noqa

Load the frames. In this example we are loading the M3CD dataset with the reduced number of stuctures.

m3cd_frames = ase.io.read("data/explore_m3cd.xyz", ":")

We are defining a function used in chemiscope.explore() as a featurizer that computes the descriptors using MACE-MP0 and then applies t-SNE. Basically, we repeat the steps done in the previous example but using different mace calculator.

def mace_mp0_tsne(frames, environments):
    if environments is not None:
        raise ValueError("'environments' are not supported by this featurizer")

    # Initialise a mace-mp0 calculator
    descriptor_opt = {"model": "small", "device": "cpu", "default_dtype": "float64"}
    calculator = mace_mp(**descriptor_opt)

    # Calculate the features
    descriptors = []
    for frame in frames:
        structure_avg = np.mean(
            (calculator.get_descriptors(frame, invariants_only=True)),
            axis=0,
        )
        descriptors.append(structure_avg)
    descriptors = np.array(descriptors)

    n_jobs = min(len(frames), os.cpu_count())

    # Apply t-SNE
    perplexity = min(30, descriptors.shape[0] - 1)
    reducer = TSNE(n_components=2, perplexity=perplexity, n_jobs=n_jobs)
    return reducer.fit_transform(descriptors)

Provide the created function to chemiscope.explore().

cs = chemiscope.explore(m3cd_frames, featurize=mace_mp0_tsne)

Downloading MACE model from 'https://tinyurl.com/46jrkm3v'
Cached MACE model to /home/runner/.cache/mace/46jrkm3v
Using Materials Project MACE for MACECalculator with /home/runner/.cache/mace/46jrkm3v
Using float64 for MACECalculator, which is slower but more accurate. Recommended for geometry optimization.
/home/runner/work/chemiscope/chemiscope/.tox/docs/lib/python3.11/site-packages/mace/calculators/mace.py:128: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  torch.load(f=model_path, map_location=device)
/home/runner/work/chemiscope/chemiscope/.tox/docs/lib/python3.11/site-packages/chemiscope/structures/__init__.py:76: UserWarning: the following atomic properties are only defined for a subset of frames: ['tags']; they will be ignored
  return _ase_list_atom_properties(frames)

To show case the result, we are loading pre-computed data using the mace_mp0_tsne function for 1k structures.

fetch_dataset("mace-mp-tsne-m3cd.json.gz")
chemiscope.show_input("data/mace-mp-tsne-m3cd.json.gz")

×

Example with SOAP, t-SNE and environments

This example demonstrates how to compute descriptors using the SOAP and t-SNE with environments parameter specifying which atoms in the frames are used for calculating the descriptors.

We are defining a custom featurizer that takes frames and environments.

def soap_tnse_with_environments(frames, environments):
    if environments is None:
        raise ValueError("'environments' must be provided")

    grouped_envs = {}
    unique_structures = set()

    # Get atom-centered indices from environments
    for [env_index, atom_index, _cutoff] in environments:
        if env_index not in grouped_envs:
            grouped_envs[env_index] = []
        grouped_envs[env_index].append(atom_index)
        unique_structures.add(env_index)
    centers = list(grouped_envs.values())

    # only include frames that are present in the environments
    if len(unique_structures) != len(frames):
        frames = [frames[index] for index in sorted(unique_structures)]

    # Get global species
    species = set()
    for frame in frames:
        species.update(frame.get_chemical_symbols())
    species = list(species)

    # Initialize calculator
    soap = SOAP(
        species=species,
        r_cut=4.5,
        n_max=8,
        l_max=6,
        sigma=0.2,
        rbf="gto",
        average="outer",
        periodic=True,
        weighting={"function": "pow", "c": 1, "m": 5, "d": 1, "r0": 3.5},
        compression={"mode": "mu1nu1"},
    )

    # Calculate descriptors
    feats = soap.create(frames, centers=centers)

    # Compute tsne
    perplexity = min(30, feats.shape[0] - 1)
    reducer = TSNE(n_components=2, perplexity=perplexity)
    return reducer.fit_transform(feats)

Provide a created function and environments to chemiscope.explore(). The environments are manually defined following the format [index of structure, index of atom, cutoff].

chemiscope.explore(
    frames=m3cd_frames,
    featurize=soap_tnse_with_environments,
    environments=[(1, 2, 3.5), (2, 0, 3.5)],
)

×

/home/runner/work/chemiscope/chemiscope/.tox/docs/lib/python3.11/site-packages/chemiscope/structures/__init__.py:76: UserWarning: the following atomic properties are only defined for a subset of frames: ['tags']; they will be ignored
  return _ase_list_atom_properties(frames)