Benchmarking Physics-Inspired Machine Learning Models for Transition Metal Complexes

This repository provides datasets and scripts for benchmarking physics-inspired machine learning models for property prediction of transition metal complexes with diverse charge and spin states.

Setup

Prerequisites

• Python 3.12 or later
• conda or pip

Dependencies The Python scripts require NumPy, Pandas, and ASE.

pip install -r requirements.txt

Data

The data/ directory contains all datasets, including XYZ files, property CSV files, extended XYZ files (for MACE), and dataset splits.
It includes three datasets: TM-GSspinPlus, tmPHOTO, and OctaKulik. See the included README.md for details.

Molecular Representations

Packages used to generate molecular representations can be built from source by following the official documentation:

• ε-SPAHM, SPAHM(a), SPAHM(b): Q-stack

  git clone https://github.com/lcmd-epfl/Q-stack
  cd Q-stack
  git checkout versioning_issue
  pip install -e .[spahm,regression]

• SLATM, FCHL: Install the modified qml2 version below, forked from qml2.
  It includes minor updates to atomic SLATM generation without affecting the resulting representations.

  git clone https://github.com/lcmd-epfl/tmc_qml2
  cd tmc_qml2
  pip install -e .

• SOAP: featomic, Documentation

  pip install featomic

Bash scripts for generating job files, along with the required Python scripts, are provided under:

• Molecular representation generation and kernel computations (including timing measurements): representations/
  See the included README.md for details.

• Kernel ridge regression using Q-stack: krr/
  See the included README.md for details.

Precomputed NumPy arrays of the molecular representations used in this work are available on Materials Cloud.

3DMol

Installation instructions and example input files for 3DMol are available at: https://github.com/lcmd-epfl/3DMol/tree/TMC-benchmark-v0
Trained models and log files are available on Materials Cloud.

MACE

Install the modified MACE version for intensive property prediction:

git clone https://github.com/lcmd-epfl/tmc_mace
cd tmc_mace
git checkout intensive
pip install -e .

Example job scripts and train/test extended XYZ files are provided in mace/ for each dataset subdirectory.
See the included README.md for details.
Trained models, job scripts, and log files are available on Materials Cloud.

Environments

Two Conda environment files are provided as references.
Both environments were built on Red Hat Enterprise Linux 9.4 (Plow), x86_64.

Benchmark Environment

The environment file environment_x86_64-rhel_9.yml includes all dependencies required for molecular representation generation and kernel benchmarking, including the following packages:

• qstack
• qml2
• featomic
• mace

conda env create -f environment_x86_64-rhel_9.yml
conda activate benchmark_tmc

MACE-Only Environment

This mace environment mace_x86_64-rhel_9.yml includes only the dependencies required for training MACE models.

conda env create -f mace_x86_64-rhel_9.yml
conda activate mace