In this Letter, we consider optical bound states in the continuum (BICs) in the infrared range supported by an all-dielectric metasurface in the form of subwavelength dielectric grating. We apply the random forest machine learning method to predict the frequency of the BICs as dependent on the optical and geometric parameters of the metasurface. It is found that the machine learning approach outperforms the standard least square method at the size of the dataset of ≈4000 specimens. It is shown that the random forest approach can be applied for predicting the subband in the infrared spectrum into which the BIC falls. The important feature parameters that affect the BIC wavelength are identified.