Phase equilibrium calculation, also known as flash calculation, has been extensively applied in petroleum engineering, not only as a standalone application for a separation process but also as an integral component of compositional reservoir simulation. Previous research devoted numerous efforts to improve the accuracy of phase equilibrium calculations, which place more importance on safety than speed. However, the equation-of-state-based flash calculation consumes an enormous amount of computational time in compositional simulation and thus becomes a bottleneck to the broad application of compositional simulators. Therefore, it is of vital importance to accelerate flash calculation without much compromise in accuracy and reliability, turning it into an active research topic in the past two decades. With the rapid development of computational techniques, machine learning brings another wave of technology innovation. As a subfield of machine learning, the deep neural network becomes a promising computational technique due to its great capacity to deal with complicated nonlinear functions, and it thus attracts increasing attention from academia and industry. In this study, we establish a deep neural network model to approximate the iterative flash calculation at given moles, volume, and temperature, known as the NVT flash. A dynamic model designed for NVT flash problems is iteratively solved to generate data for training the neural network. In order to test the model’s capacity to handle complex fluid mixtures, three real reservoir fluids are investigated, including one Bakken oil and two Eagle Ford oils. Compared to previous studies that follow the conventional flash framework in which stability testing precedes phase splitting calculation, we incorporate stability test and phase split calculation together and accomplish two steps by a single deep learning model. The trained model is able to identify the single vapor, single liquid, and vapor–liquid states under the subcritical region of the hydrocarbon mixtures. A number of examples are presented to show the accuracy and efficiency of the proposed deep neural network. It is found that the trained model makes predictions at most 244 times faster than the iterative NVT flash calculation for the given cases and meanwhile preserves high accuracy.