Feat: lematrho, populate forces, stress_tensor, energy from static DFT runs. - #50
Feat: lematrho, populate forces, stress_tensor, energy from static DFT runs.#50speckhard wants to merge 2 commits into
Conversation
…sprun - Add vasprun.xml.gz to STATIC_FILES so it downloads in the same loop as charge density files; remove the now-redundant separate relax vasprun download - Replace parse_vasprun_structure (returned only Structure) with parse_vasprun_output (returns structure, forces, stress_tensor, energy) by parsing ionic_steps[-1] and vasprun.final_energy - Wire forces, stress_tensor, and energy into the OptimadeStructure constructor; energy_corrected is auto-computed (equals energy for R2SCAN since the MP 2020 correction is PBE-only) - Fix functional from Functional.PBE to Functional.r2SCAN (confirmed from INCAR: METAGGA = R2scan) - Update all test patches and mock return values accordingly; add TestVasprunForces covering parse_vasprun_output and end-to-end pass-through of forces/stress/energy into the Parquet row
eaa67e8 to
926a715
Compare
|
Just realized that the entire pipeline only extracts the last structure of all the trajectories. Is that also only what the final dataset contains? If that's a case, would a LeMatRho-Traj dataset be too expensive to store? I think it would also be interesting to have this dataset I can take care of writing a PR if the data is already available in the S3 bucket to support this. |
Ramlaoui
left a comment
Ramlaoui
left a comment
There was a problem hiding this comment.
Thanks @speckhard this looks great!
| xx, yy, zz, xy, yz, xz = (float(x) for x in s.flat[:6]) | ||
| stress_out = [[xx, xy, xz], [xy, yy, yz], [xz, yz, zz]] | ||
|
|
||
| try: |
There was a problem hiding this comment.
Silent failure, do we not have energy and forces on all structures?
| forces_out = np.asarray(frc, dtype=float).reshape(-1, 3).tolist() | ||
|
|
||
| strs = final.get("stress") # Stress tensor in 3 × 3 (kBar) | ||
| if strs is not None: |
There was a problem hiding this comment.
Seems surprising that vasp outputs might have different shapes for stress, shouldn't we only expect one in that case. Also from what I remember, vasp stress is in kBar and ML datasets are in eV/A^3, LeMat-Bulk also keeps kBar but will be worth mentioning in the dataset card to avoid confusion.
| # Determine which raw files to keep | ||
| need_raw = set() | ||
| # Determine which raw files to keep (use .gz names, matching STATIC_FILES). | ||
| # vasprun.xml.gz is always needed for structure, forces, stress, and energy. |
There was a problem hiding this comment.
Isn't that a required file for every structure or can it not be present in the bucket?
| if tool_paths["can_run_ddec6"]: | ||
| need_raw |= _DDEC6_FILES | ||
|
|
||
| for filename in STATIC_FILES: |
There was a problem hiding this comment.
Nitpick: the branching would be simpler if the loop was done on all the files except for vasprun.xml.gz which would be handled outside and not be part of STATIC_FILES. But not important at all.
After talking to @mfranckel, the reason for this is clear now structures are already in equilibrium because they come from LeMat-Bulk :). |
Hmm good point, but we relax with R2SCAN structures from LeMatBulk that were often relaxed with PBE so maybe we should save the trajectory even a few points of minimally different forces/energies would be useful |
2 participants
Summary
LeMatRhoStaticMaker(NSW=0,already-relaxed geometry) instead of
LeMatRhoRelaxMaker_1. Forces/stressfrom the static run are the physically meaningful residual forces at the
relaxed geometry.
forces(nsites×3, eV/Å),stress_tensor(3×3, kBar),energy(eV), andenergy_corrected(eV — equalsenergyfor R2SCAN sincethe MP 2020 correction is PBE-only).
Functional.PBEtoFunctional.r2SCAN,confirmed from INCAR files in S3 (
METAGGA = R2scan).vasprun.xml.gzmerged into the existingSTATIC_FILESloop, removing a separate relax-vasprun download step.
Test plan
.env.integration)functional=r2scan, non-nullforces,stress_tensor,energy,energy_corrected == energyTestVasprunForcesadded coveringparse_vasprun_outputand end-to-endpass-through into the Parquet row