What is the secret of learning organic chemistry

Unravel the secrets of chemical bonding with machine learning

A new approach to machine learning offers important insights into catalysis, a fundamental process that makes it possible to reduce the emission of toxic fumes or to create important materials such as tissue.

In an in Nature Communications published report, Hongliang Xin, Associate Professor of Chemical Engineering at Virginia Tech, and his research team developed a Bayesian learning model of chemisorption, or Bayeschem for short, with the aim of using artificial intelligence to decipher the nature of chemical bonding to catalyst surfaces.

"It all depends on how catalysts bind to molecules," said Xin. "The interaction must be strong enough to break some chemical bonds at relatively low temperatures, but not so strong that the catalysts would be poisoned by reaction intermediates. This rule is known as the Sabatier principle in catalysis."

Understanding how catalysts interact with various intermediates and determining how to control their bond strengths so that they are within this "goldilocks zone" is key to designing efficient catalytic processes, Xin said. Research provides a tool for this purpose.

Bayeschem works with Bayesian learning, a special machine learning algorithm for deriving models from data. "Let's say you have a domain model that is based on well-established physical laws and you want to use it to make predictions or learn something new about the world," said Siwen Wang, a former PhD student in chemical engineering. "The Bayesian approach consists of learning the distribution of the model parameters in light of our previous knowledge and the often sparse observed data, while at the same time enabling the uncertainty quantification of model predictions.

The d-band theory of chemisorption used in Bayeschem is a theory that describes chemical bonding to solid surfaces with d electrons, which are usually shaped like a four-leaf clover. The model explains how d orbitals of catalyst atoms overlap and are attracted to adsorbate valence orbitals, which have a spherical or dumbbell-like shape. It has been considered the standard model in heterogeneous catalysis since its development by Hammer and Nørskov in the 1990s, and although it has been successful in explaining binding trends for many systems, the model occasionally fails due to the inherent complexity of electronic interactions, according to Xin.

According to Xin, Bayeschem takes D-band theory to a new level in order to quantify these interaction strengths and possibly adjust some buttons, such as structure and composition, to design better materials. The approach advances the d-band theory of chemisorption by expanding its predictive and interpretive capabilities of adsorption properties, both of which are critical to the discovery of catalysts. However, when compared to the black box machine learning models trained on large amounts of data, Bayeschem's predictive accuracy has room for improvement, said Hemanth Pillai, a chemical engineering graduate student in Xin's group who also participated in the study .

"The ability to develop highly precise and interpretable models based on in-depth learning algorithms and chemisorption theory is very rewarding for achieving the goals of artificial intelligence in catalysis," said Xin.