Long-Term Creep Curve Interpolation of TP321H Stainless Steel Using a Neural Network Trained on a Limited Set of Experimental Data

Long-term creep testing is expensive and time-consuming, which makes a reliable creep curve prediction essential for the design and safety assessment of high-temperature components. This paper presents a data-driven approach using a backpropagation neural network (BPNN) to interpolate the full creep curve from a limited set of experimental data. Creep tests were conducted on TP321H stainless steel at 550 °C under five stress levels (160–200 MPa) for up to 8000 h. A three-hidden-layer BPNN was trained on only four stress levels (160, 170, 190, and 200 MPa) and subsequently applied to predict the complete creep curve in the held-out 180 MPa condition. Its predictive performance was compared with the classical θ-projection method. The BPNN prediction agrees closely with the experimental curve, with R2 = 0.990 and RMSE = 0.0547%, compared with 0.770 and 0.1443% for the θ-projection method. These results indicate that, for the present material and test conditions, a BPNN trained on four creep curves can accurately interpolate the creep curve at 180 MPa without constitutive assumptions.