As oil and gas exploration is advanced to deeper and ultra-deep layers, drilling efficiency problems are increasingly highlighted. Mechanical specific energy theory is used to evaluate drilling efficiency. In deep drilling, energy losses are caused by drill string friction and power drilling tools. Significant errors are found in mechanical specific energy calculations based on surface data. A bidirectional recurrent neural network model was developed for downhole drill bit motion state prediction. Various influencing factors were considered in the model. The relationship between surface parameters (depth, drilling pressure, rotary table speed, torque, and displacement) and downhole measurement data was learned. Accurate predictions of drill bit pressure, torque, and speed were achieved. The experimental results indicate that the model’s average absolute error is 0.05.The root mean square error is 0.07.The explained variance score is 0.91.The coefficient of determination is 0.92.Case analysis shows that the average relative error is only 2.3%. This error is significantly lower than the 49.7% of traditional methods. Compared with the benchmark models such as recurrent neural network, gated recurrent unit, long and short-term memory network and gradient boosting tree, the proposed two-way recurrent neural network further highlights the advantages of the model by reducing the average absolute error and root mean square error by 9%~15% and 4%~10%, respectively. A transfer learning optimization strategy was proposed for model adaptability between different blocks. The model’s understanding of basic physical drilling characteristics was retained. Only 15% of target block data is needed for effective model transfer compared to source block training data. A feasible solution is provided for large-scale application of intelligent mechanical specific energy real-time calculation technology. This technology is applicable in deep and ultra-deep oil and gas field development.