Auto Machine Learning Assisted Preparation of Carboxylic Acid by TEMPO‐Catalyzed Primary Alcohol Oxidation

Abstract

Comprehensive Summary Though alcohol oxidations were considered as well‐established reactions, selecting productive conditions or predicting reaction yields for unseen alcohols remained as major challenges. Herein, an auto machine learning (ML) model for TEMPO‐catalyzed oxidation of primary alcohols to the corresponding carboxylic acids is disclosed. A dataset of 3444 data, consisting of 282 primary alcohols and 45 conditions, were generated using high‐throughput experimentation (HTE). With the HTE data and 105 descriptors, a multi‐label prediction was performed with AutoGluon (an open‐source auto machine learning framework) and KNIME (an open‐source data analytics platform). For the independent test of 240 reactions (a full matrix of 20 unseen alcohols and 12 conditions), AutoGluon with multi‐label prediction for yield prediction (AGMP) gave excellent performance. For external test of 1308 reactions (consisting of 84 alcohols and 45 conditions), AGMP still afforded good results with R2 as 0.767 and MAE as 4.9%. The model also revealed that the newly generated descriptor (Y/N, classification of the reaction reactivity) was the most relevant descriptor for yield prediction, offering a new perspective to integrate HTE and ML in organic synthesis.

HTE and machine learning-assisted development of iridium(i)-catalyzed selective O–H bond insertion reactions toward carboxymethyl ketones

“…Notably, a novel and simple descriptor called molecular additive fingerprint (MAF), 14 developed by our group, was used in this study. 15 The results showed that MAF was more effective than other descriptors in terms of reaction prediction (for details, see the ESI, Table S13†). With the MAF descriptor, 4 ML methods, 16 including Support Vector Regression (SVR), Gradient Boosted Trees (GBT), Random Forest Regression (RF), and eXtreme Gradient Boosting (XGB), were evaluated.…”

XMe ‐ Xiamen Molecular Electronics Code: An Intelligent and Open‐Source Data Analysis Tool for Single‐Molecule Conductance Measurements†

“…On the other hand, the non‐linear, non‐stationary, and high‐dimensional nature of the conductance data imposes further complexity on the development of data mining approaches. Fortunately, machine learning methods that have achieved extraordinary success in processing big data in different fields such as computer vision, speech recognition, and neural language processing, and also in the field of chemistry, [ 52‐56 ] show promising potential in understanding the conductance data. [ 17,22‐23 ] Clustering analysis is one such machine learning method that is aimed at revealing the intrinsic structure of the dataset that cannot be identified directly or through primary statistical approaches.…”

HTE- and AI-assisted development of DHP-catalyzed decarboxylative selenation

“…Feature description has a crucial impact on the model performance. In this work, a simple descriptor called molecular additive fingerprint (MAF), derived from extended-connectivity fingerprints (ECFPs) 16 and developed by our group recently, 17 was found to be effective for reaction prediction. MAF was calculated via RDKit.…”