Training machine learning potentials for reactive systems: A Colab tutorial on basic models.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry

Background:

• Machine learning potential (MLP) models are increasingly used for molecular systems.
• Training system-specific MLPs for reactive systems is crucial for accelerating simulations.
• Existing methods require accessible training resources for new researchers.

Purpose of the Study:

• To provide a self-guided Colab tutorial for training system-specific MLPs for reactive systems.
• To familiarize researchers with fundamental techniques for free energy simulations.
• To support the broader research community in utilizing MLPs for chemical and enzyme reactions.

Main Methods:

• Introduction to feedforward neural networks (FNN) and Gaussian process regression (GPR) models.
• Utilizing symmetry functions (including ANI) and embedding neural networks (DeepPot-SE) as molecular descriptors.
• Applying FNN and GPR models with extracted features to reproduce energies and forces for reactive molecular configurations.

Main Results:

• Demonstration of fitting the Müller-Brown potential using FNN and GPR models.
• Successful feature extraction using symmetry functions and embedding neural networks.
• Reproduction of energies and forces for a Claisen rearrangement reaction using trained MLPs.

Conclusions:

• The Colab tutorial offers a practical approach to learning MLP model training for reactive systems.
• The presented methods facilitate the acceleration of free energy simulations for chemical reactions.
• This resource aims to empower young researchers in computational chemistry and related fields.