Understanding the mechanical behavior of granular materials is of paramount importance in various geotechnical applications. Coral sand, a naturally occurring sediment composed of broken coral fragments, plays a crucial role in marine engineering. However, characterizing and predicting the breakage behavior remain challenging due to its complex and heterogeneous nature. At current work, a set of one-dimensional compression tests were carried out considering varying initial conditions. Four machine learning (ML) algorithms (random forest, linear regression, fully connected neural network, and eXtreme Gradient Boosting) were adopted to predict particle breakage ratio. The initial loading stress, fines content, density state, grain size, coefficient of uniformity, and curvature coefficient were considered as variables for regression and classification. Relative particle breakage ratio was set as output feature. The dataset was divided into 25 and 75% as the test and training sets, respectively. Test results show that high stress, lower fines content, smaller relative density, and larger grain can lead to remarkable particle breakage. ML analysis suggests that both random forest and eXtreme Gradient Boosting achieved remarkable accuracy levels, exceeding 99%. However, linear regression, with a root mean squared error of 0.041, presented poor performance for particle breakage prediction. The developed approach can be used to evaluate the particle breakage with an acceptable breakage modeling accuracy.