LLMSkillHub
mirroredCopied from beita6969/ScienceClaw under MIT. beita6969 has not claimed this page, and is not affiliated with LLM SkillHub.

beita6969 / scienceclaw-benchmark-for51

Use when a task asks to predict a phonon property of inorganic crystals from their structure, for example the frequency of the last peak of the phonon density of states in cm^-1 (Matbench phonons), given structures (lattice, species, fractional coordinates) and a labelled training set; the deliverable is a float vector in input order, judged by mean absolute error. Covers composition/geometry descriptors, log-target tree regressors and pretrained universal interatomic potentials (SevenNet, CHGNet) as phonon feature extractors. Corresponds to FoR51 of the companion ScienceClaw-Eval benchmark.

0
SKILL.md
3.4 KB

Phonon frequency from crystal structure

Task

  • Inputs: structures {formula, lattice (3x3 row vectors, angstrom), species, frac_coords}; labelled training structures with the target in cm^-1; structures to predict.
  • Deliverable: a 1-D float array, one value per structure in input order, in cm^-1 (1 THz = 33.356 cm^-1; convert explicitly if a tool returns THz).

Quality

Mean absolute error in cm^-1, lower is better. A sensible baseline is 5-nearest-neighbour regression on standardised composition/geometry descriptors.

Tools

  • Descriptors are built in a code node (no scilib function): site-weighted statistics of atomic number, mass, electronegativity, covalent radius, group, period and valence electrons, plus n_sites, volume per atom, density, packing fraction and nearest-neighbour distances.
  • scilib.matphonon_regression.fit_predict(X_train, y_train, X_eval, ...): ExtraTrees on the log target, three seeds averaged, clipped to 0..5000 cm^-1 (typed operator phonon_frequency_regression); fit_hist_predict is a HistGradientBoosting alternative when trees over-smooth (phonon_frequency_regression_boosted); cross_validate(X, y, folds=5) gives mae (phonon_regression_cv_mae). X must be finite and y positive; return_info=True adds provenance.
  • scilib.matphonon_mlip.phonon_features(structures, mesh=8, min_len=7.0, disp=0.01, model="sevennet"): phonopy finite displacements on the unrelaxed structure with a pretrained universal potential (model is sevennet, sevennet_mf0_pbe, sevennet_mf0_r2scan or chgnet); returns {"X": (n, 21), "names": [...]} of frequency statistics and DOS peaks in cm^-1, NaN rows for failed structures (operators phonon_feature_matrix_mlip, phonon_features_mlip). Check scilib.matphonon_mlip.available() and scienceclaw_tools(operation=status); needs sevenn or chgnet, plus phonopy and torch; weights ship with the pip packages (asset ids sevennet, chgnet). About 0.25 s (CHGNet) to 0.9 s (SevenNet) per structure on a GPU. Concatenate these columns with the descriptors and fit the regressor.

Rules

  • Universal potentials soften frequencies systematically: use the MLIP columns as regression inputs, never as predictions.
  • Fit only on labelled training rows; impute or drop NaN feature rows with training statistics only.
  • SevenNet-l3i5 and CHGNet were trained on Materials Project trajectories (energies, forces, stresses, not phonon frequencies), which can include the task crystals; disclose that overlap.
  • Shuffled K-fold is optimistic when new crystals contain elements absent from training; estimate under shift with leave-element-out folds.
  • Keep unit and shape; predictions finite and positive. SevenNet on CPU repeats only to about 1e-5 relative, so set SCIENCECLAW_MLIP_CACHE for exact replay.