jwchen25/smgp

MatInvent — Retrieval-Retro / rr_thermo assets

Model weights and reference data for the retrieval_retro and rr_thermo reward calculators in MatInvent. These are too large for git (~1.6 GB) and are fetched at runtime with hf_hub_download.

Files are byte-identical between the two calculators, so one copy serves both, except MP_Energetics.json, which only rr_thermo uses.

random_mpc_pool.pt and embed/mpc_embeddings_pool.pt concatenate the original train/valid/test splits in that order (28434 rows: 0–22746 / 22747–25589 / 25590–28433). The order is load-bearing — train_dataset[i] must stay aligned with row i of the embedding matrix.

Source

Retrieval-Retro — https://github.com/HeewoongNoh/Retrieval-Retro

Dataset downloaded from the link in that repository; the two *_pool files are our merge of its train/valid/test splits.

• random_mpc_pool.pt (1.3G) — retrieval pool
• mpc_embeddings_pool.pt (3.5M) — retrieval embeddings
• element_embeddings.json (432K) — element vocabulary
• random_template.json (7.4K) — index → formula map
• random_precursor_formation_energy.pt (4.5K) — precursor formation energies

Retrieval-Retro: Retrieval-based Inorganic Retrosynthesis with Expert Knowledge. NeurIPS 2024.

Trained here

Trained on the dataset above using the Retrieval-Retro code. Hyperparameters are encoded in the filenames (split, batch size, learning rate, hidden size, seed).

• Retrieval_Retro_128_ours_Retrieval_Retro_random_3_best.pt (46M) — main model
• TL_pretrain(formation_exp)_embedder(graphnetwork)_lr(0.0005)_batch_size(256)_hidden(256)_seed(0)_.pt (17M) — NRE formation-energy model
• mpc_best.pt (373K) — MPC retriever

ARROWS — https://github.com/njszym/ARROWS

Redistributed under MIT (© Nathan J. Szymanski). Contents are Materials Project formation energies (Bartel-corrected), CC-BY 4.0.

• ARROWS/arrows/energetics/MP_Energetics.json (216M) — read at import time

N. J. Szymanski et al. Nature Communications (2023). https://doi.org/10.1038/s41467-023-42329-9