Machine Learning & Neural Network Potentials

Mon, 01 Jan 2024 00:00:00 +0000

Overview

Machine learning is transforming computational chemistry, and we are harnessing it in two complementary ways:

IQA-Informed Neural Network Potentials

Classical machine-learned interatomic potentials (NNPs) are trained on total energies and forces, but lack chemical interpretability. We develop NNPs informed by IQA energy components (self-energies and interaction energies), resulting in potentials that:

Decompose into physically meaningful atomic and pairwise contributions.
Transfer more reliably to out-of-distribution chemical environments.
Naturally encode the correct physics of bonding interactions.

Topological Descriptors as ML Features

QTAIM atomic properties and IQA energy components serve as physically motivated features for machine learning models targeting molecular properties, reaction barriers, and drug–target binding affinities.

Deep Learning for Electron Density

We explore the use of deep learning models to predict electron densities directly, enabling rapid computation of topological properties for large molecular datasets.

Codes & Tools

Our ML work builds on open-source frameworks (PyTorch, JAX) and is integrated with our in-house topological analysis codes.

MM2SF

Mon, 01 Jan 2024 00:00:00 +0000

About

MM2SF automatically generates optimised Atom-Centred Symmetry Functions (ACSFs) for use as descriptors in neural network interatomic potentials. Given a molecular dynamics trajectory or normal-mode sampling, it applies a Gaussian Mixture Model (GMM) to decompose the chemical space into well-defined clusters, then selects symmetry function parameters that accurately describe each region.

Supported symmetry function types:

Two-body (radial) — G^rad
Three-body (angular) — G^ang (modified functional form)

Installation

pip install git+https://github.com/m-gallegos/MM2SF.git

Or from a downloaded zip:

pip install MM2SF-main.zip

Neural Network Potentials | QTCOVI – Theoretical and Computational Chemistry