The solution of atmospheric chemical kinetics is considered as the dominant computationally intensive task in chemical transport models. Reported strategies for improving computational efficiency have not changed the essence of solving ordinary differential equations. In this study, we emulate a typical gas-phase chemistry solver implemented in chemical transport models in Chinese Mainland using a 28-layer residual regression neural network. Validation results have a total R² of 0.985 with the original numerical solution for 194 species during a single timestep. Multiple-timestep test for one month verified the accuracy and stability of our model while no exponential accumulated bias was observed, though a slow accumulation of biases existed. There were 102 species with R² values higher than 0.90 and 137 species with daily NMB values below 30% among the 194 species throughout the month. Meanwhile, the computational efficiency of gas-phase chemistry using the deep learning emulator was promoted by 10.6 times on one CPU and 85.2 times on one GPU. Our study confirms the feasibility of emulating an atmospheric chemical kinetic solver with data-driven models through deep learning and represents a step towards a deep learning solver being used in a chemical transport model. The deep learning emulator is expected to not only accelerate atmospheric chemical kinetic solutions by orders of magnitude but also provide a promising pathway for implementing near-explicit rather than lumped chemical mechanisms in chemical transport models.