Skip to content

DMF_P3_RUNTIME_VALIDATION_2026-06-18.md

Latest commit

 

History

History
97 lines (76 loc) · 4.33 KB

DMF_P3_RUNTIME_VALIDATION_2026-06-18.md

File metadata and controls

97 lines (76 loc) · 4.33 KB
Raw

DMF P3 Runtime Validation 2026-06-18

Version: 2026-06-18 Date: 2026-06-18 Status: Candidate-comparison evidence; candidate energy/force write fixed after job 525521 Owner: ATST-Tools maintainers Scope: P3 Direct MaxFlux candidate runs on SAI with DP backend

Summary

Two staged DMF candidate-generation cases completed on SAI Slurm and were compared with existing ATST ABACUS and DP reference results:

  • 01_neb_Li-Si
  • 02_neb_H2-Au

This is P3 candidate-comparison evidence, not production TS validation. DMF remains experimental because the generated candidates have not yet been refined with Dimer/Sella/CCQN or validated by vibration/IRC.

Runtime

Field Value
Slurm job 525521
State COMPLETED, exit 0:0
Partition/node 4V100, 4v100n32
Resources --nodes=1, --ntasks=1, --gpus-per-node=2, --qos=rush-1o2gpu
Elapsed 00:02:41
MaxRSS 14632788K
Environment bootstrap Node-local clone of dpeva-dpa4-test, offline cached cyipopt/IPOPT install, editable ATST checkout install with --no-build-isolation --no-deps
Command path examples/18_dmf_production_validation/submit_dmf_dp_rush_1gpu.sbatch

Two failed setup attempts preceded the successful run:

  • 524936: one-GPU bootstrap was killed during conda clone/transaction.
  • 525006: two-GPU bootstrap passed clone but compute-node conda could not reach conda-forge.
  • 525153: offline conda install passed, but pip build isolation tried to reach PyPI.
  • 525187 and 525200: DMF reached candidate generation, then exposed JSON serialization issues for NumPy arrays and bytes in IPOPT info; fixed by the DMF summary normalizer.

These failures are now covered by the current sbatch script and tests/unit/test_dmf_workflow.py.

Candidate Comparison

The completed job generated dmf_validation_report.json with schema atst-dmf-validation-suite-v1, status candidate_compared, and two cases.

2026-06-21 correction note: the table below is pre-fix standalone evidence. It shows 7 images because the ATST wrapper did not forward YAML nmove=10 into the final DirectMaxFlux solve, so PyDMF used its own default nmove=5. The wrapper now forwards nmove and writes final t_eval to dmf_summary.json; the post-fix D2S rough-DMF rerun in job 533228 recorded nmove=10, n_images=12, and t_eval length 12 for both staged cases. A standalone rerun attempt, job 533227, failed during compute-node conda environment setup before ATST execution (tk package cache file missing), so this report keeps the original standalone numeric comparison as historical evidence only.

Case tmax Images IPOPT status RMSD to ABACUS TS (A) RMSD to DP TS (A) ABACUS barrier (eV) DP barrier (eV)
01_neb_Li-Si 0.4994841198 7 success 0.0128141221 0.0126187834 0.618346 0.7231818847
02_neb_H2-Au 0.5554593655 7 success 0.0551754077 0.0253945726 1.124752 0.6474094018

Both cases used:

  • initial_path: linear
  • pbc_mode: cartesian_unwrapped
  • remove_rotation_and_translation: false
  • confirm_pbc_risk: true
  • DP model temp_repos/dp_model/DPA-3.1-3M.pt, head Omat24

Interpretation

The two candidates are geometrically close to existing reference transition state structures under same-order Cartesian RMSD, especially for Li-Si. H2-Au is also close to the DP reference and within 0.06 A of the ABACUS reference.

The original job 525521 showed that candidate trajectories did not retain single-point energy/force results in the written dmf_tmax.traj, so candidate energy and force fields in that generated comparison report were null/not finite. The workflow now evaluates the tmax candidate once before writing it and records candidate energy/fmax in dmf_summary.json; this fix is covered by tests/unit/test_dmf_workflow.py.

Remaining P3/P4 Work

  • Add or run a refinement stage from DMF tmax into Dimer/Sella/CCQN for at least the two completed cases.
  • Run vibration/IRC validation on refined candidates and record TS mode, endpoint connection, walltime, and failure modes.
  • d2s.rough_method: dmf now has unit-level runtime integration and remains experimental. Do not mark it supported until the refinement and vibration/IRC evidence above is available.