Abstract: Fatigue crack growth tests were conducted on TA2 commercial pure titanium under different load ratios （0.1–0.8）, load amplitudes （1 441–1 882 N） and stress intensity factor amplitudes （7.3–50.2 MPa·m^0.5）. With the test data, a fatigue crack propagation rate model was established by the backpropagation neural network optimized by genetic algorithm （GA-BPNN） method. The prediction accuracy and error of the GA-BPNN model under single feature variable of stress intensity factor amplitude, double feature variables of load ratio and stress intensity factor amplitude, three feature variables of load ratio, stress intensity factor amplitude and load amplitude were analyzed and compared with results of the BPNN model and the traditional Forman equation. The results show that the fatigue crack growth rate in the early growth stage showed a linear characteristic with the stress intensity factor amplitude, while it showed a nonlinear characteristic in the later stage of crack growth. Under high load amplitudes, the linear growth stage of fatigue cracks was short and the fatigue crack growth rate was low. With increasing load ratio, the difference in fatigue crack growth rates under high and low load amplitudes became larger. With the increase of the number of feature variables, the prediction accuracy of the GA-BPNN model for the fatigue crack growth rate decreased. The prediction accuracy of the GA-BPNN model was higher than those of the BPNN model and the traditional Forman formula. With the increase of the number of feature variables, the prediction advantage of the GA-BPNN model became more significant, and its prediction effect was better under low load ratios of 0.1 and 0.5. With the increase of the load ratio, the prediction error of the GA-BPNN model increased. Under the same load ratio, the prediction error under the higher load amplitude was smaller.