Experimental Validation of Calibration Constants Determined by a Deep Operator Network for IHD in Laminated Composite Materials

Abstract Background The incremental hole-drilling technique requires calibration constants to relate the measured strain data to the residual stress distribution that existed in the material prior to drilling. For laminated composite materials, the calibration constants must be determined individually from finite element (FE) analyses for the specific combination of laminate configuration, material properties and experimental parameters used. Depending on the computational method and uncertainty sources considered, hundreds of FE analyses may be required per laminate configuration. This can become burdensome, especially if numerous laminate configurations are under investigation. A type of neural network known as a deep operator network (DeepONet) offers the potential to significantly reduce the amount of computation required. Objective Investigate the use of a DeepONet to predict calibration constants for the integral method applied to balanced and symmetric fibre reinforced plastic (FRP) laminates of eight plies aligned at 0 $$^\circ $$ ∘ , $$90^\circ $$ 90 ∘ , +45 $$^\circ $$ ∘ or -45 $$^\circ $$ ∘ . Methods The DeepONet model was trained on the FE predicted top surface displacement fields, for every stress application depth and drilling increment, of 15 of the 70 possible E-glass/epoxy FRP laminate configurations. IHD was performed on two specimens, one each of [90 2 /0 2 ] s and [0 2 /+45/-45] s configuration, which were not part of the training set. Results The residual stress distributions obtained using DeepONet predicted and FE generated calibration constants showed good agreement, $$\approx $$ ≈ 1 – 1.5 MPa, over the measurement depth. Conclusions The agreement in the calculated residual stress distributions demonstrates that the calibration constants predicted by the DeepONet model have acceptable accuracy for practical application, $$\approx $$ ≈ 10%, while significantly reducing the amount of FE analyses required.