Machine Learning Studies of Organic Reactions Catalyzed by Nickel Catalysts with Nitrogen-Containing Ligands

Nitrogen ligands is one of the key components in metal-catalyzed organic synthesis. The steric and electronic properties of nitrogen ligands enables modification of reaction activity, selectivity, and catalyst stability. Despite massive empirical knowledge from experiments, a good design of improved nitrogen ligands for specific organic reactions remains challenging due to the difficulties and complexity of the underlying chemical space. Meanwhile, machine learning methods is emerging as a powerful tool for uncovering structure–reactivity relationships from different sources of experimental and computational data, offering a possible route toward data-driven catalyst optimization. This thesis explores organic reactions catalysed by nickel catalyst along with nitrogen ligands through a machine learning approaches. Ligands and reaction substrates are represented using a combination of different descriptors and molecular fingerprints. These descriptors are essentially evaluated in terms of their ability to encode relevant chemical information while remaining recognizable for machines. Then, a wild range of datasets for selected nitrogen-ligand-mediated reactions, especially reductive cross-coupling reactions, are used to train a regression and classification model with the techniques of neural networks. Model performance is assessed with respect to predictive accuracy, robustness to limited and noisy data, and transferability across related reaction spaces. The analyses of prediction result is conducted by the interpretation of different datapoints to identify the key attributes that control the reaction yields, rates, and enantioselectivities. Finally, the evaluated models preserve the possibility to be applied in a designated context to prioritize new ligand candidates and be predicted under different conditions. Further Comparisons with literature data and mechanistic insights can illustrate how machine learning models are able to complement traditional catalyst development strategies.