An efficient inverse method for nuclear graphite damage based on physics-informed neural networks

As a critical structural material in High-Temperature Gas-cooled Reactors (HTGRs), nuclear graphite undergoes mechanical property degradation that affects structural integrity and operational safety. Consequently, accurately and efficiently characterizing its damage evolution law is imperative. However, inverse methods based on finite element models suffer from computational inefficiency, model dependency, and insufficient convergence stability. To address these limitations, this paper proposes a physically consistent damage parameter inverse method based on Physics-Informed Neural Network (PINN) for unirradiated nuclear graphite under room-temperature conditions, enabling efficient inverse analysis of key damage parameters. Our method treats the damage parameters as trainable variables and constructs a multi-term loss function combining strain-field error with physical constraints, including momentum conservation, moment equilibrium, and axial force equilibrium. In four-point bending tests with simulated nuclear-graphite data, the proposed method significantly outperforms Finite Element Model Updating (FEMU) in both parameter inversion and strain-field prediction, reducing prediction errors by more than 50%. Further validation using experimental Digital Image Correlation (DIC) nuclear-graphite data demonstrates that PINN maintains low prediction error and reliable identification. Additional tests on an aluminum alloy specimen with known parameters confirm feasibility. Moreover, the proposed framework reduces the computational time for parameter identification by more than 99% compared with FEMU.