Abstarct: Rock mass drillability is a quantitative index which is widely used in geology, mining engineering, petroleum engineering, and other engineering disciplines for evaluating how hard or easy is a rock mass to drill. Due to paramount importance of this parameter in improving the results of the models used to predict rate of penetration (ROP) of drilling tools, a wide spectrum of research has been done on determination of this parameter within the scope of mining and civil engineering. This is while, laboratory measurement of this parameter has been addressed in only limited number of oil and gas research works. In addition, due to failure to account for real well conditions, the introduced rock mass drillability index (RDI) in these research works is not practically applicable. Therefore, the present research develops a new RDI for oil and gas reservoir rocks which is consistent with the corresponding ROP model. For this purpose, operating parameters, petrophysical logs, and drilling reports from four wells drilled into three different fields were collected. Then, in order to determine geomechanical parameters of rock mass, one-dimensional well modeling was performed. In order to undertake extensive analyses on the collected data and understand the value of rock mass parameters in ROP models, the first level of data including three different databaxses was formed. The first databaxse was introduced by operating parameters of each well only, while second and third databaxses were developed baxsed on not only the operating parameters, but also petrophysical logs and geomechanical properties of rock mass, respectively. Since the results of field data-baxsed models were highly dependent on the acquisition quality, raw data were subjected to preprocessing to attenuate the effect of noise and select the most significant parameters affecting ROP. In this respect, first, using spectral analysis and autocorrelation assessments, the presence and intensity of noise within data was assessed. Then, Savitzky–Golay (SG) filter was applied to attenuate the effect of noise on useful data. Continuing with the research, in order to select superior features in ROP estimation, plus-l take-r search method was adopted on the first-level databaxses. On order to establish identical conditions for analyzing the results of ROP, five parameters in each databaxse were considered for estimating ROP. Subsequently, ROP was modeled using multivariate nonlinear regression and multilxayer perceptron (MLP) neural network methods baxsed on the selected parameters in each databaxse only. An investigation on the obtained values of ROP from the models developed baxsed upon different first-level databaxses indicated that, on average, accuracies of the ROP estimation models developed baxsed on the databaxses 2 and 3 using the regression method exceeded that of the model developed baxsed on the databaxse 1 by 7% and 13%, respectively. The same result was found for the models prepared using the neural network method, with the corresponding figures being 13% and 18%, respectively. Given the importance of rock properties in ROP models, continuing with the study, a number of punch tests were performed on core samples recovered from similar formations along two wells to extract rock mass brittleness and drillability. The force-penetration curve obtained for each rock sample in this test was used to determine three brittleness indices as well as the newly developed index called rock mass drillability index. Investigation of the relationships between each of these indexes and ROP showed that, the developed index is well related to ROP via an exponential relationship. As such, in the next step, petrophysical and rock geomechanical logs were used to present a relationship for estimating RDI. Results of this evaluation demonstrated that, one can estimate the newly developed RDI at good accuracy using two geomechanical properties of rock mass, namely confined compressive strength and internal friction angle. Therefore, this index was calculated at all of the studied wells and the second-level databaxses containing operating parameters along with the proposed drillability index for each well were formed. In order to validate the new RDI in ROP models, Non-dominated Sorting Genetic Algorithm (NSGA) was implemented on the second-level data for feature selection. The use of selected second-level parameters at each well in ROP modeling showed that, on average, accuracies of regression and MLP NN models were higher than that of the corresponding models baxsed on the third first-level databaxse by 16% and 28%, respectively. Moreover, a comparison between the results of the regression-baxsed ROP estimator developed upon second-level data with the ROP model proposed by Burgin and Yang, when applied on the studied wells, showed that, on average, the second-level models were about 27% more accurate than the Burgin and Yang’s. This is while, the model proposed by Burgin and Yang uses 9 different parameters while the proposed models in this research use as low as five parameters only.