🚀 ABOUT

This repo contains Supporting Materials for the article "Challenges and Opportunities for Machine Learning Potentials in Transition Path Sampling: Alanine Dipeptide and Azobenzene Studies" by Nikita Fedik, Wei Li, Nicholas Lubbers, Benjamin Nebgen, Sergei Tretiak, Ying Wai Li

📄 Final article on publisher site - OPEN ACCESS

If you have any comments/questions/requests, feel free to contact me:
✉️ [email protected]
🌐 nikitafedik.xyz

---

ML interatomic potentials

HIP-NN-TS potential:
hippynn - official github repo of HIP-NN-TS code
HIP-NN - original paper
HIP-NN-TS [tensor-sensitive] - new paper

HIP-NN-TS trained to ANI-1x in TPS article
best model used for all results except active learning section: models/hipnnts_data_ani1x_only_before_al_seed533257

HIP-NN-TS trained to ANI-1x + 10800 structures from TPS trajectories in active learning section
path to model: models/hipnnts_data_ani1x_tps_10800_dataseed0_modelseed<x>
path to training script: models/hippnn_ani_al0.py
docs for loading pretrained model

ANI-1X potential
TorchAni - github.com/aiqm/torchani
TorchAni - paper
ANI - original paper

Note

All ANI-1x calculations in the article were done by vanilla TorchANI with parameters defined in https://github.com/aiqm/ani-model-zoo/tree/master/resources/ani-1ccx_8x

---

⌨️ SCRIPTS

scripts contains all scripts mainly as Jupyter notebooks to reproduce the results in the article:

• run thermal MD for AD and AZ
• evaluate E and F using HIP-NN-TS (hippynn), ANI-1x (torchani), Amber 14 (through GAFF) and Sage by OpenFF
• get accuracy metrics, plot correlation plots and get coverage of phase space for both AD and AZ

Prerequisites:

• numpy
• torch
• natsorted
• torchani
• hippynn
• OpenMM
• OpenFF
• openbabel

It's best to create a dedicated conda environment for this project.

---

DATA

ANI-1x database
download ANI-1x dataset (5.21 GB)
ANI-1x - "Less is more" paper

Note

All data for the project is available at Zenodo. Just clone the repo and download the data in the root, e.g. ml_tps_si:
git clone https://github.com/nikitafedik/ml_tps_si.git
cd ml_tps_si
wget https://zenodo.org/records/15047941/files/data.tar.zst
tar --use-compress-program=zstd -xf data.tar.zst

alanine dipeptide (AD)
10k snapshots from MD trajectories
data/AD/thermal_MD_10k/DFT-logs - DFT logs for 10k selected snapshots data/AD/thermal_MD_10k/ - npy arrays with E and F and xyz of important conformers

12k points most visited configurations from TPS trajectories

data/AD/10800_seed0-<x>.npy- train data
data/AD/1200_seed0-<x>.npy- test data where x:

• E_formatioin_QM_kcal_mol = formation energies (full DFT E - E of all atoms) | kcal/mol
• G_QM_kcal_mol_A = gradients (not forces) | kcal/mol/A
• R = atomic positions | Angstrom
• Z = atomic numbers corresponsing to coordinates in R array

azobenzene (AZ)
10k snapshots from MD trajectories
data/AZ/thermal_MD_10k/DFT-logs - DFT logs for 10k selected snapshots data/AZ/thermal_MD_10k/ - npy arrays with E and F and xyz of important conformers

full isomzerization path at DFT/UDFT levels
path is a concatenation of (reopt of last + IRC to trans-AD + TS + IRC to cis-AD + reopt of last IRC point)
see data/AZ/<x>/DFT-logs for more information or visualization

data/AZ/cs-inversion/AD_cs-DFT_inversion_path.xyz - inversion path | closed-shell DFT
data/AZ/os-rotationAD_os-DFT_rotation_path.xyz - rotation path | open-shell DFT

⚙️ OTHER TOOLS:

OpenPathSampling - Python package for TPS
TPS - comprehensive review
ASE - Atomic Simulation Environment