reproduction/: the alignment model trained with seven seeds (checkpoints, results, checksums)

reproduction/README.md

@@ -0,0 +1,23 @@
+# reproduction/ — the alignment model, trained with seven seeds
+
+Seven checkpoints of `tri_curri_v1_save.py`, the alignment training script of MEIDNet, trained with seeds 0 to 6 on all
+18,928 Perov-5 materials (2,200 epochs, Python 3.10, torch 2.6.0, one NVIDIA A100, about 7 hours each).
+
+`results.json` gives, per seed, the cosine similarity and the L2 distance between the structure latent and the property
+latent of the same material (normalised encoder outputs, before the projection heads) and the structure-matching rate.
+`SHA256SUMS.txt` lists the checksums.
+
+| | cosine | L2 | structure matching |
+|---|---|---|---|
+| mean ± s.d. over the seven seeds | 0.954 ± 0.011 | 0.301 ± 0.034 | 86.1 ± 5.0 % |
+
+Load a checkpoint with `meidnet.checkpoint.load_checkpoint(path)` (version-1 format, ABX₃ perovskites, properties
+`heat_all` and `dir_gap`), or recompute the numbers on a laptop in about a minute:
+
+```bash
+pip install git+https://github.com/ABnano/MEIDNet.git
+meidnet download-data
+python scripts/reproduce_alignment.py --seed 4
+```
+
+Results, figures and analysis: https://babu09-meidnet.hf.space/docs/benchmarks/perov5-reproduction.html

reproduction/SHA256SUMS.txt

@@ -0,0 +1,9 @@
+9af6307788fcdbd7994c2d18aa085c5ce03c43c2003cef740e8197a891095b08 *meidnet_paper_rerun_seed0.pth
+67f0b0cbc9e58eae24aff693d79d0c15c9df975b22c18f4203ea5e0ce5bdc2ce *meidnet_paper_rerun_seed1.pth
+3e82d459ccb517056f701fd1ab5cd6f215dd7eceefe25fc574b2f561e3115c02 *meidnet_paper_rerun_seed2.pth
+205de8b833900c54841d9dcfe3510a8e4a8a4bb9bbf5516d80cb68a7fc121889 *meidnet_paper_rerun_seed3.pth
+5da9b4dc68b3e98d7c60ff9ba0eef6e062bb54d66d4b1d8cda78857dc50b36ee *meidnet_paper_rerun_seed4.pth
+3f70e752a6f7aeb0e0df05d898a91c54bfd33b246c129cdfda50eba8265f066a *meidnet_paper_rerun_seed5.pth
+399a17f23ec9a1c1be73eeed88c10b5e8b4004f9c77e712cb82b7c45f3c1fe38 *meidnet_paper_rerun_seed6.pth
+95b26aa9660bf641e72089c1d91cbd3ae63ae89f14ce9072adf573629e9f3bf4 *results.json
+51c7a9958c2f285d15c65a3b955e98310f20ac86870a9b9b746c0719de88fd0a *README.md

reproduction/meidnet_paper_rerun_seed0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9af6307788fcdbd7994c2d18aa085c5ce03c43c2003cef740e8197a891095b08
+size 2838994

reproduction/meidnet_paper_rerun_seed1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:67f0b0cbc9e58eae24aff693d79d0c15c9df975b22c18f4203ea5e0ce5bdc2ce
+size 2838994

reproduction/meidnet_paper_rerun_seed2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e82d459ccb517056f701fd1ab5cd6f215dd7eceefe25fc574b2f561e3115c02
+size 2838994

reproduction/meidnet_paper_rerun_seed3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:205de8b833900c54841d9dcfe3510a8e4a8a4bb9bbf5516d80cb68a7fc121889
+size 2838994

reproduction/meidnet_paper_rerun_seed4.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5da9b4dc68b3e98d7c60ff9ba0eef6e062bb54d66d4b1d8cda78857dc50b36ee
+size 2838994

reproduction/meidnet_paper_rerun_seed5.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3f70e752a6f7aeb0e0df05d898a91c54bfd33b246c129cdfda50eba8265f066a
+size 2838994

reproduction/meidnet_paper_rerun_seed6.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:399a17f23ec9a1c1be73eeed88c10b5e8b4004f9c77e712cb82b7c45f3c1fe38
+size 2838538

reproduction/results.json

@@ -0,0 +1,68 @@
+{
+ "what": "tri_curri_v1_save.py (the paper's training script, unchanged), retrained 2026-10-04 on all 18,928 Perov-5 materials; cosine and L2 between the normalised encoder outputs (before the projection heads), on the same materials",
+ "runs": [
+  {
+   "file": "meidnet_paper_rerun_seed0.pth",
+   "seed": 0,
+   "cosine": 0.9563,
+   "l2": 0.2934,
+   "structure_matching_pct": 90.04,
+   "epochs": 2200,
+   "script": "tri_curri_v1_save.py"
+  },
+  {
+   "file": "meidnet_paper_rerun_seed1.pth",
+   "seed": 1,
+   "cosine": 0.9585,
+   "l2": 0.2862,
+   "structure_matching_pct": 91.41,
+   "epochs": 2200,
+   "script": "tri_curri_v1_save.py"
+  },
+  {
+   "file": "meidnet_paper_rerun_seed2.pth",
+   "seed": 2,
+   "cosine": 0.9373,
+   "l2": 0.3529,
+   "structure_matching_pct": 82.91,
+   "epochs": 2200,
+   "script": "tri_curri_v1_save.py"
+  },
+  {
+   "file": "meidnet_paper_rerun_seed3.pth",
+   "seed": 3,
+   "cosine": 0.9592,
+   "l2": 0.2839,
+   "structure_matching_pct": 91.55,
+   "epochs": 2200,
+   "script": "tri_curri_v1_save.py"
+  },
+  {
+   "file": "meidnet_paper_rerun_seed4.pth",
+   "seed": 4,
+   "cosine": 0.9654,
+   "l2": 0.2615,
+   "structure_matching_pct": 78.21,
+   "epochs": 2200,
+   "script": "tri_curri_v1_save.py"
+  },
+  {
+   "file": "meidnet_paper_rerun_seed5.pth",
+   "seed": 5,
+   "cosine": 0.9586,
+   "l2": 0.2863,
+   "structure_matching_pct": 85.15,
+   "epochs": 2200,
+   "script": "tri_curri_v1_save.py"
+  },
+  {
+   "file": "meidnet_paper_rerun_seed6.pth",
+   "seed": 6,
+   "cosine": 0.94,
+   "l2": 0.3449,
+   "structure_matching_pct": 83.76,
+   "epochs": 2200,
+   "script": "tri_curri_v1_save.py"
+  }
+ ]
+}