ANN Models Of Flow Stress Of Hot Deformation And Microstructure Evolution In Hydrogenized BT20 Alloy

The hot deformation behavior and microstructure evolution of BT20 titamium alloy hydrogenated were studied and artificial neural network (ANN) models were established to predict them in this paper.All specimens were hydrogenised at 800℃for 10 min~2 hours to obtain different hydrogen content in the range of 0~0.8 wt%. Hot compression experiments were conducted at deformation temperature 600, 700 and 800℃, original strain rate 8.3×10-2, 8.3×10-3 and 8.3×10-4s-1, settled height reduction 60%. After that, drawing of ture strain-ture stress curves and measurement of microstructure parameters were going on.It is found that hydrogenation is able to improve the plasticity of BT20 alloy and lower the flow stress during the hot deformation. With different deformation parameters, the soft mechanism of alloy includes phase transformation, dynamic recovery and dynamic recrystallization.With experiment data, two parts of simulation model have been built by using BP neural network. First part had only one model for predicting flow stress whose inputs were: time of hydrogenization t, original strain rateε& , deformation temperature T, true strainεand hydrogen content CH, output flow stressσ. In the end, the numerical results gained via the networks were compared with the experiment results. It appears that the agreement is reasonably good, the mean relative error within 10% and the mean absolute error below 15 MPa. Also, the simulation of smooth curve is better than that of non-smooth curve.Second part had four models for microstructure evolution during the deformation process. The inputs of each model were t,ε& , T, CH, and outputs involved primaryαvolume fraction Vα, grain length l, aspect ratioγ. There were three models with only one of these outputs individually, and one with all of them combined. The mean relative errors of 4 models are not more than 10%, which show that the predicted results are in satisfactory agreement with the experiment. However, the model with single output is more accurate than that with combined outputs.