AbstractIn the era of automation and artificial intelligence, some problems in the petroleum industry can be efficiently solved by machine-learning models based on big data. Previous researchers have addressed truly precise methods to evaluate reserves from accessible well-test data, although enormous time, effort, and cost were required. However, due to the lack of sufficient conventional material balance (CMB) test data for a large number of wells in the Su Dong (SD) gas field, only 12 wells (0.99%) were generally suitable for previous methods. We substituted the fully built-up reservoir pressures used in CMB with those converted from shut-in casing-pressure data. Suffering from data noise caused by pressure fluctuation, only 22.91% of the plots were linear. The nonlinear plots were enhanced into linear plots by denoising and smoothening according to engineering knowledge. Machine-learning models were developed to perform the enhancement automatically, which constructed an adaptive and feasible workflow. The model improved the percentage of applicable wells to 68.16% (824 out of 1,209) and achieved considerable precision for reserve evaluation. The proposed method is much more suitable than CMB in gas fields featuring unstable production status and provides feasible reserve evaluation. This paper presents an exemplary demonstration of the combination of petroleum engineering and machine learning.