Convolutional Neural Networks for Estimation of Uniaxial Tensile Test Equivalent Properties from Small Punch Test

The Small Punch Test (SPT) has been developed as a small sample technique for the evaluation of mechanical properties of structural materials. However, the SPT is subject to systematic biases, resulting in inaccurate estimates of Uniaxial Tensile Test (UTT) properties. In this study, an experimental approach has been adopted to investigate the potential of neural networks to predict UTT-equivalent behavior from SPT measurements. An experimental database containing paired SPT and UTT data has been prepared for three boiler steels (10H2M, 13HMF, and 15HM) in both new and service-degraded states. Convolutional neural networks (CNN) have been trained and evaluated for curve-to-curve prediction. The working hypothesis is that CNN models, by exploiting local curve features, are capable of reducing the systematic bias inherent to SPT, generating estimates of UTT properties with precision comparable to conventional UTT measurements. Consistent with trends in applied deep learning, the results confirm the robustness of convolutional architectures. In general, the findings provide strong evidence that CNNs can translate SPT data into UTT equivalent material properties, thereby bridging a long-standing methodological gap and supporting automated evaluation of structural steels in service.