Systems and techniques are described for data encoding using a machine learning approach to generate a distortion prediction {circumflex over (D)} and a predicted bit rate {circumflex over (R)}, and to use {circumflex over (D)} and {circumflex over (R)} to perform rate-distortion optimization (RDO). For example, a video encoder can generate the distortion prediction {circumflex over (D)} and the bit rate residual prediction based on outputs of the one or more neural networks in response to the one or more neural networks receiving a residual portion of a block of a video frame as input. The video encoder can determine bit rate metadata prediction based on metadata associated with a mode of compression, and determine {circumflex over (R)} to be the sum of and . The video encoder can determine a rate-distortion cost prediction Ĵ as a function of {circumflex over (D)} and {circumflex over (R)}, and can determine a prediction mode for compressing the block based on Ĵ.