[Data Pipeline] Request for CIF-to-npz preprocessing script

[breatheccc]

[Data Pipeline] Request for CIF-to-npz preprocessing script — missing link in the mobility prediction pipeline

First of all, thank you for the excellent work on OCNet! The paper (npj Computational Materials, 2025, 11, 308) presents a very promising framework for multi-scale virtual characterization of organic functional materials.

---

What I found

I am trying to reproduce the full CIF → electron mobility pipeline described in the README:

cd ./biomolecular_properties/code/film_ll_scripts_elec
 && python lmdb_convert.py mol_105511_mob && bash infer.sh
 && python mobility_film.py mol_105901_mob OCNet

After thorough investigation, I discovered that the zhengcheng branch contains the three downstream scripts referenced above — lmdb_convert.py, infer.sh, and mobility_film.py — which handle the LMDB conversion → OCNet inference → kMC mobility stages. Thank you for making these available!

However, lmdb_convert.py (line 57–63) reads a pre-computed data_ll.npz file:

data = np.load(input_path, allow_pickle=True)
coord = data['coord']        # dimer 3D coordinates
symbol = data['symbol']      # atom types
features = data['feature']   # 8-dimensional TB-level descriptors
mol_name = data['name']      # molecule identifiers

No script in either the main or zhengcheng branches, nor in any of the four Zenodo records (14935486, 14934728, 14934618, 15083880), generates these data_ll.npz files from raw CIF or dimer structures. The .gitignore excludes biomolecular_properties/code/*_scripts, which suggests the preprocessing scripts exist in the development environment but were not committed or distributed.

Additionally, I noticed that mobility_film.py (line 3) references gen_sol_box.py — likely another preprocessing utility that is also not present in the repository.

---

Primary Request

Could you please share the preprocessing script(s) that generate data_ll.npz files from CIF or film structures?

According to the paper, this step uses xTB to compute the 8-dimensional tight-binding-level electronic descriptors (overlap integrals, orbital-specific effective transfer integrals, and total effective transfer integrals). Specifically, the script should handle:

1. Extracting molecular dimers from the crystal/film structure (center-of-mass distance < 10 Å)
2. Computing the 3D coordinates and atom types for each dimer
3. Running xTB to obtain TB-level electronic descriptors → 8-dimensional feature vector
4. Computing the reorganization energy (reorg_e), displacement vectors (displace), and dimer pair indices (pair)
5. Saving everything in the .npz format expected by lmdb_convert.py

Any code snippet, configuration file, or even a brief description of the workflow would be tremendously helpful for reproducibility.

---

Additional Technical Questions

To ensure full compatibility with the pre-trained weights, I would appreciate clarification on a few implementation details:

1. n_max (maximum number of atoms). Is this 1024 (as suggested by --max-atoms default in trans_mix_feature_infer.py:80), or a different value used during pre-training?

2. Vocabulary mapping. The pretrain_dict.txt file lists 33 elements (including [PAD], [CLS], [SEP], [UNK]). Is this the exact element-to-index mapping used during pre-training?

3. Gaussian kernel parameters a_ij and b_ij. The GaussianLayer in unimol.py uses learned nn.Embedding layers for mul and bias per edge type. Are the pre-trained values of these embeddings accessible, or should they be (1, 0) for inference?

4. PAD masking strategy. Is the padding mask computed simply as src_tokens.eq(self.padding_idx) (as in unimol.py:205), or is there an additional attention mask generation step?

---

Acknowledgment

I will properly cite OCNet (doi: 10.1038/s41524-025-01788-y) in any subsequent academic work. Thank you for your time and consideration!