This review discusses the application of machine learning in laser-induced breakdown spectroscopy (LIBS). LIBS is a powerful analytical technique with advantages such as simple sample preparation, fast detection, and minimal sample damage. However, its performance is limited by the spatial inhomogeneity and temporal variability of the laser-induced plasma, leading to poor quantification results from traditional models. Machine learning offers a promising solution by establishing multivariate regression models that can better handle signal fluctuations and matrix effects, thus improving both qualitative and quantitative performance. The review summarizes the progress of machine learning in LIBS from two main aspects: data preprocessing and machine learning methods. Data preprocessing includes spectral selection, variable reconstruction, and denoising to enhance performance. Machine learning methods aim to improve quantification performance while reducing the impact of matrix effects and spectral fluctuations. The review also highlights challenges in future research, such as limitations in training data, the gap between physical principles and algorithms, and the need for better generalization and data processing capabilities. Machine learning methods are categorized into unsupervised, supervised, and semi-supervised learning. Unsupervised learning identifies patterns in unlabeled data, supervised learning uses labeled data to build predictive models, and semi-supervised learning combines both. These methods are applied to improve data preprocessing and analysis in LIBS. The review emphasizes the role of machine learning in enhancing analysis repeatability and suppressing matrix effects. It also proposes application prospects and suggestions for future research in LIBS using machine learning. The review concludes that integrating physical principles with machine learning could be a key solution for improving LIBS quantification, especially when large datasets are available.This review discusses the application of machine learning in laser-induced breakdown spectroscopy (LIBS). LIBS is a powerful analytical technique with advantages such as simple sample preparation, fast detection, and minimal sample damage. However, its performance is limited by the spatial inhomogeneity and temporal variability of the laser-induced plasma, leading to poor quantification results from traditional models. Machine learning offers a promising solution by establishing multivariate regression models that can better handle signal fluctuations and matrix effects, thus improving both qualitative and quantitative performance. The review summarizes the progress of machine learning in LIBS from two main aspects: data preprocessing and machine learning methods. Data preprocessing includes spectral selection, variable reconstruction, and denoising to enhance performance. Machine learning methods aim to improve quantification performance while reducing the impact of matrix effects and spectral fluctuations. The review also highlights challenges in future research, such as limitations in training data, the gap between physical principles and algorithms, and the need for better generalization and data processing capabilities. Machine learning methods are categorized into unsupervised, supervised, and semi-supervised learning. Unsupervised learning identifies patterns in unlabeled data, supervised learning uses labeled data to build predictive models, and semi-supervised learning combines both. These methods are applied to improve data preprocessing and analysis in LIBS. The review emphasizes the role of machine learning in enhancing analysis repeatability and suppressing matrix effects. It also proposes application prospects and suggestions for future research in LIBS using machine learning. The review concludes that integrating physical principles with machine learning could be a key solution for improving LIBS quantification, especially when large datasets are available.