Machine Learning Modeling of Environmentally Relevant Chemical Reactions for Organic Compounds

Abstract

Environmental chemical reactions have been frequently investigated for various purposes; however, it remains challenging to accurately model either the reaction kinetics or reaction pathways. Existing studies mostly model reaction kinetics with traditional quantitative structure−activity relationships (QSARs) or reaction pathways with reaction template methods; however, these approaches generally require extensive feature engineering or manual extraction of reaction templates. Recently, machine learning (ML) has become a promising tool for modeling chemical reactions as ML models can perform well and are powerful in using diverse chemical representations. This Review starts with a concise comparison of traditional and ML modeling approaches for chemical reactions, followed by a brief discussion of the status of and future needs in modeling environmental organic reactions. Data collection and data cleaning techniques for reaction kinetics and pathways are then discussed. We then summarize the advantages and limitations of commonly used chemical representations and feature selection techniques. Next, we critically review general ML model evaluation and interpretation processes and propose a three-step evaluation process, that is, comparisons with general metrics, baseline models, and existing models. Lastly, we explore ML modeling approaches for small data sets, including transfer learning and active learning, which have been successfully employed in many other fields, for future modeling of environmental chemical reactions.

Predicting Rate Constants of Alkane Cracking Reactions Using Machine Learning

“…Recently, machine learning, which forms the basis of artificial intelligence and data science, has shown great potential to predict reaction rate constants. − In this regard, some groups have devoted their efforts to developing machine learning based models to predict the activation energies and minimum-energy paths of chemical reactions. − For example, Lewis-Atwell et al highlighted the formidable capability of machine learning in predicting activation energies of chemical reactions, particularly with neural network models . The models utilizing geometric features as inputs were found to offer the advantage of lower computational costs, while those that incorporate quantum chemical features improved predictive performance .…”

Structure-Based Reaction Descriptors for Predicting Rate Constants by Machine Learning: Application to Hydrogen Abstraction from Alkanes by CH3/H/O Radicals

“…Recently, machine learning, lying at the core of artificial intelligence and data science, has emerged as a promising method to yield highly reliable reaction rate constants. − The machine learning methods can be in principle clarified into three categories: supervised learning, unsupervised learning, and reinforcement learning, in which the supervised machine learning is usually applied in predicting chemical reaction properties by using different molecular representations as inputs. , In this regard, many pioneering works have been performed to learn activation energies and minimum energy paths of chemical reactions. ,− Meanwhile, some efforts have been made to directly predict rate constants …”

“…The absolute relative error of the predicted logarithmic rate constants, defined as the absolute value of the difference between the experimental and predicted rate constants divided by the experimental value, was less than 4%. Zhong et al combined a convolutional neural network with molecular image to model radical rate constants of water contaminants, whose predictive performance was comparable to the molecular fingerprint-deep neural network model developed by the same group. ,− Greaves et al predicted rate constants for organic processes in mixtures containing ionic liquids by applying multiple linear regression and artificial neural networks with descriptors taken mostly from the Dragon descriptor database . To make the established machine learning models broadly available, Sanches-Neto et al developed a web application “pySiRC” that predicts reaction rate constants of radical-based oxidation processes of aqueous organic contaminants by combining three machine learning algorithms with molecular fingerprints. , …”

Advancing the Comprehension of Iron Ions with Different Valences in Anammox-Based Processes: Insight into the Response and Mechanism Based on Prediction and Classification Machine Learning Models

“…It is a regulation method based on mathematical analysis and reaction mechanisms to conduct numerical fitting and prediction . Machine learning has been applied to environmental problems in fields such as biochar and organic chemistry. − The application of machine learning to predict and classify environmental problems is a new breakthrough, but this technology is rarely used in anammox-based systems. Meanwhile, iron ion is unstable, resulting in iron-containing wastewater being a more complicated system.…”