Ru/H introduction

Trained model: M3GNet, starting from the Universal Potential (UP)

Reference method: PBE-D3(BJ) with engine Quantum ESPRESSO

System: H atoms depositing onto Ru surfaces

../../../_images/ams.gif

Problem: M3GNet-UP-2022 is not very reliable for high-temperature surface chemistry.

Solution: Retraining the model gives better agreement.

Expected duration: This example takes several days to run on a modern compute node.

This is a very thorough example which shows how to

  • construct initial reference data using PES Scans like volume scans, cartesian coordinate scans, and bond scans as well as MD simulations

  • training an initial model to the reference data before the active learning loop

  • running an active learning loop with the molecule gun

Important

This tutorial is only compatible with AMS2024.102 or later. AMS2024.101 will not work.

To run this example, you may download common_ru_h.py, 01_Ru_volume_scan_H2_bond_scan.py, 02_surface_pes_scans.py, 03_Ru_H2_gas_snapshots.py, 04_Ru_H_initial_training.py, 05_active_learning_molecule_gun_md.py and run

#!/bin/sh

# also make sure that common_ru_h.py is in the current directory

"$AMSBIN/amspython" 01_Ru_volume_scan_H2_bond_scan.py || exit 1
"$AMSBIN/amspython" 02_surface_pes_scans.py || exit 1
"$AMSBIN/amspython" 03_Ru_H2_gas_snapshots.py || exit 1
"$AMSBIN/amspython" 04_Ru_H_initial_training.py || exit 1
"$AMSBIN/amspython" 05_active_learning_molecule_gun_md.py || exit 1

Important

The common_ru_h.py file contains a variable TESTING_MODE.

Set TESTING_MODE = True to not use DFT reference calculations but instead a custom-trained M3GNet model for the reference calculations. This will let you run through the workflow quickly without running any expensive DFT reference calculations.

Alternatively, you can follow the individual Jupyter notebooks below: