checkpoints + per-cell metadata 2026-09-06T11:41:13-04:00 (e957a54)

f52a701 verified about 1 month ago

2.76 kB

license: mit
tags:
  - materials-science
  - crystal-structure-prediction
  - diffusion
  - alignn
  - ablation
library_name: alignn

ALIGNN-CSP angular-diffusion ablations — checkpoints

Trained checkpoints for every cell of the line-graph × bond-angle-diffusion ablation of ALIGNN-CSP (a conditional diffusion model for crystal structure prediction), plus two partial checkpoints from the earlier A-suite's cancelled Alexandria port. Source, harness, analysis and full provenance: https://github.com/crhysc/alignn-csp-ablation (read PROJECT_STATE.md there).

Layout

lgmatrix/<dataset>/<arm>_seed0/
  best_model.pt          EMA weights at the epoch of minimum validation structural loss (what was benchmarked)
  config.json            every hyperparameter the run was trained with, incl. angle_mode and n_parameters
  history.json           per-epoch train/val losses (total, lattice, frac, angle, structural)
  metrics_sym.json       AtomBench metrics, symmetrised post-relaxation pipeline
  metrics_nosym.json     AtomBench metrics, unsymmetrised post-relaxation pipeline
  generation_config.json exact generate_benchmark.py arguments
  ABLATION.yaml          the exhaustive per-run record (what it is, how it differs, jobs, hardware, sha256s)
angle-ablation/alex/<arm>_seed0/   partial (cancelled at epoch <=375) A0 and A1, for the record only

<dataset> is jarvis (JARVIS-DFT Supercon-3D, 847/105/103) or alex (Alexandria DS-A/DS-B, 6603/825/825). <arm>:

arm	angular tier	angle_mode	line graph
nolg	none	off	no (9 pair convs)
A0	none	off	yes (3 ALIGNN + 3 pair)
nolg_ad	derived (legacy)	derived_aux	no
A3	derived (legacy)	derived_aux	yes
nolg_b3	independent	independent	no
B3	independent	independent	yes

Loading

from alignn.inverse.sample import load_model
model, schedule, normalizer, cfg = load_model("lgmatrix/alex/B3_seed0/best_model.pt", device="cuda", use_ema=True)

with the alignn package from https://github.com/crhysc/alignn at branch lg-angle-diffusion-matrix (commit f8121f4 or later; the independent cells need that branch).

What these were trained with

AdamW, lr 1e-3 one-cycle, batch 64, hidden 256, T=1000, σ∈[0.005,0.5], cosine ᾱ, loss weights (lattice, frac, angle) = (1, 10, 1), EMA 0.999, seed 0, checkpoint selected on validation structural loss. Full protocol and the reading of the results: EXPERIMENT_SET.yaml in the harness directory of the GitHub repository.

Caveat

Every benchmark number shipped beside these weights was scored after ALIGNN-FF relaxation of 32 candidates. The generator-alone scores are the project's open task (see NEXT_TASK.md in the repository).