Considering the highly nonlinear behavior and phenomenological complexity of critical heat flux (CHF), this study proposes a novel method to predict CHF on microstructure surface using machine learning technologies. An extensive literature survey was conducted to collect experimental data on microstructure surfaces. Data on horizontal silicon specimens of cylindrical pillars with square arrangements were selected for both training and testing various machine learning methods, including ν-support vector machine, back-propagation neural network, radial basis function neural network, general regression neural network, and deep belief network (DBN). Among the tested machine learning methods, DBN is shown to provide the best accuracy for CHF prediction. The obtained parametric CHF behavior of DBN with respect to pillar diameter, spacing, and height agrees with the physical understanding of CHF on microstructure surfaces. The presented approach is expected to support the design optimization of microstructure for CHF maximization.