Conversation
Parameter sets: - Add SLAKONET_MODELS registry mapping slakonet_v0/v1/v1a to their Figshare file IDs, replacing the hardcoded v0 download URL. - default_model(model_name=...), --model_path <name> and the SLAKONET_MODEL env var now select a set; load_trained_model resolves a bare registry name through the download/cache path. Forces / stress / energy (all validated against finite differences): - Geometry stores lengths in Bohr, so dE/dR is eV/Bohr; convert to eV/Angstrom instead of applying the beta=0.1 fudge factor. - Fix the stress: dE/dh^T @ h (not dE/dh @ h^T), correct virial sign, and convert eV/Bohr^3 -> eV/Angstrom^3. The ASE wrapper also reshaped the Voigt-6 vector to (2,3) and left it in GPa, which made get_stress() raise. - alpha scaled the band energy but not the repulsive term, so the total energy was not E_band + E_rep. Default alpha and beta to 1.0; Si's equilibrium lattice constant improves from 5.72 to 5.51 A. Eigenvalues and bandgaps are unaffected. Analysis: - Add compute_site_projected_dos, layer_resolved_dos and gap_states_metric for per-atom-index projections, as needed for surface, defect and interface electronic structure.
slakonet/optim.py enabled torch.use_deterministic_algorithms(True), set_detect_anomaly(True) and set_anomaly_enabled(True) at module import. These are process-global, expensive, and were never opted into. Consequences, measured on a 2-atom Si cell with a 4x4x4 mesh (energy + forces, slakonet_v1a): anomaly + create_graph (before) 1.93 s create_graph=False 0.86 s anomaly off as well 0.32 s Anomaly mode records a Python stack trace per autograd node; profiling showed torch.fx.traceback.format_stack accounted for 84% of runtime. Worse, the global determinism flag broke force evaluation: SlakoNet's backward uses CuBLAS routines with no deterministic kernel, so autograd raised and the bare `except RuntimeError` substituted torch.zeros_like(). Any optimizer then reported instant convergence with nothing moved. It also broke unrelated calculators sharing the process - running alignn_ff and SlakoNet in one chipsff run made SlakoNet return exactly zero forces, silently. Changes: - Move the flags into an opt-in set_debug_mode(); honour SLAKONET_DEBUG. - Add create_graph (default False) to SimpleDftb. Only force-matching training needs a graph through the gradients. - Wrap SimpleDftb.calculate in allow_nondeterministic() so SlakoNet still works if another library enables determinism, restoring the caller's setting afterwards. - Never substitute zero forces on failure; raise with a diagnostic. Forces and stress are unchanged (bit-identical) and still match finite differences; 3C-SiC bandgap unchanged at 2.0995 eV.
A single fixed kpoints_array cannot serve cells that range from 2-atom bulk to defect supercells to slabs. On ideal diamond Si, where symmetry requires zero forces, the residual is: 1x1x1 2.373 eV/Ang 2x2x2 0.449 4x4x4 0.042 6x6x6 0.005 so a mesh chosen for a 64-atom supercell leaves spurious forces an order of magnitude above a typical fmax=0.05 relaxation criterion. kpoints_for() now derives the mesh from the reciprocal cell, n_i = ceil(|b_i| / kspacing), with a single k-point along non-periodic directions. kspacing=0.30 gives [7,7,7] for primitive Si, [2,2,2] for a 64-atom supercell and [4,4,1] for a slab, dropping Fmax to 0.0016. kpoints_array is still honoured when kspacing is None.
DEFAULT_MODEL_NAME kept slakonet_v1 from the upstream rebase; v1a is the current parameter set. SLAKONET_MODEL still overrides it.
SlakoNetCalculator (main.py) had only get_bandgap/get_fermi_energy; SlaKoNetCalculator (ase_calc.py) had band_structure/dos but no HS accessor. Both now expose the same three methods, with get_bandstructure/get_dos on ase_calc as aliases of the existing band_structure/dos. get_HS returns (H, S) reshaped to (n_kpoints, n_orbitals, n_orbitals); SimpleDftb stores them as (batch, n_orb, n_orb, n_k), which is awkward to index per k-point. H is in Hartree, not eV, and the basis is non-orthogonal, so the docstrings spell out the generalized eigenproblem needed to recover band energies. Verified: eigh(H[k], S[k]) * 27.211 minus E_F reproduces the calculator's eigenvalues exactly, and H and S are Hermitian at every k. Also hoist the FilteredModel wrapper out of _run_calc_with_filtered_skfs to module scope so the new methods can reuse it.
The test suite had no calculator coverage at all. test_ase_calculators adds 11 tests over both calculator classes, checking forces and stress against finite differences of the model's own energy rather than only that the calls return something -- that is the check that catches the unit and sign errors fixed earlier. Also covers get_dos, get_bandstructure, get_HS (shape, hermiticity, and that eigh(H,S)*27.211 - E_F reproduces the eigenvalues), kspacing mesh selection, agreement between the two calculator classes, and SKF re-filtering when the elements change. README: the 'Using Pretrained Models in Python' example raised KeyError: 'band_gap_eV' -- the keys are 'bandgap' and 'fermi_energy', and both are tensors. generate_shell_dict_upto_Z65 also needs model=. Document get_HS, including that H is in Hartree and needs the generalized eigenproblem. All four README python blocks now execute.
README: condense the abstract to the key claims, drop the Applications section and the Output Properties duplication, and add the dense-vs- sparse scaling table. Key Features said 'up to 2000 atoms', which the table contradicts; now says >10,000 with the sparse solver. Installation instructions kept as they were. Formatting: black -l 79 over the files touched in this branch. flake8 with the CI selection (E9,F63,F7,F82) reports 0; no line over 79 chars was introduced.
Replace the plain 'Open in Colab' text link with the standard colab-badge.svg image wrapped in the same link, so the badge itself opens the notebook.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.