This repository contains the Supporting Information code for:

Sanocki, M., & Zavadlav, J. (2025)
Generalization of Long-Range Machine Learning Potentials in Complex Chemical Spaces
Digital Discovery, vol. 5, RSC, 2026, pp. 2195–2208, DOI: 10.1039/D5DD00570A

---

📦 Repository Structure

The repository is organized into the following components:

• chemtrain/ – Training and evaluation code based on chemtrain
  → Installation: https://github.com/tummfm/chemtrain

• les/ – Training and evaluation code based on LES and MACE
  → Installation: https://github.com/ChengUCB/les/tree/main

• Dataset_creation/ – Scripts for dataset construction and biased split generation

Each submodule contains its own dependency requirements.

---

📊 Datasets

The datasets used in this study are available at:

• QMOF
  https://github.com/Andrew-S-Rosen/QMOF

• ODAC25
  https://huggingface.co/facebook/ODAC25

• OMol25
  https://huggingface.co/facebook/OMol25

---

🏋️ Training Overview

• QMOF models: use predefined dataset splits (handled internally in scripts)
• OMol25 / ODAC25 models: require split specification via command-line arguments

Split types include:

• Cluster split
• Maximum separation (maxsep)
• Random split
• Small/Large (SL) split

---

🔬 chemtrain

ODAC / OMol Training

Run chemtrain-based models as:

python SCRIPT.py GPU_ID --traj PATH_TO_TRAJ [split options]

Split Strategies

Cluster / Max Separation (external indices)

--train_indices path/to/train_indices.txt
--test_indices  path/to/test_indices.txt

Example

python allegro_efa_odac.py "1" \
  --traj path/to/dataset.traj \
  --train_indices path/to/train_indices.txt \
  --test_indices path/to/test_indices.txt

Random Split

python allegro_efa_odac.py "1" \
  --traj path/to/dataset.traj \
  --seed 3

Small/Large (SL) Split

python allegro_efa_odac.py "1" \
  --traj path/to/dataset.traj \
  --seed 3 \
  --sl True

---

QMOF Training

For QMOF, dataset splitting is handled inside each training script.

Standard Preprocessing

preprocess_mof_data(
    data_path,
    split_method="cluster",
    val_ratio=0.1,
    seed=3
)

Supported split_method values:

• cluster
• maxsep
• None (fallback to random split)

Required Files for Cluster / MaxSep

train_refcodes_<split_method>.csv
test_refcodes_<split_method>.csv

Small/Large (SL) Preprocessing

preprocess_mof_data(
    data_path,
    train_cutoff=100
)

Running Models

python SCRIPT.py GPU_IDS

---

⚙️ les

LES-based models are trained using the provided fit.sh script.

Training Entry Point

bash fit.sh

This script wraps:

run_train.py

and defines all model and training parameters.

---

📂 Dataset Format

Datasets must be provided as .extxyz files:

--train_file path/to/train.extxyz
--valid_file path/to/valid.extxyz
--test_file  path/to/test.extxyz

---

⚙️ Configuration

All key parameters are defined directly in fit.sh, including:

• Model architecture (MACELES)
• Number of channels
• Cutoff radius
• Number of interactions
• Optimization settings
• EMA / SWA

---

📈 Outputs

Each training run produces:

• Standard LES/MACE logs
• Model checkpoints
• Outputs saved in the working directory