Synergizing machine learning and multiscale shakedown method for shakedown loading capacity evaluation of parameterized lattice structures

In recent years, advanced soft computing methodologies have emerged as more effective than traditional approaches in estimating fatigue properties. However, a significant research gap remains in efficiently and accurately evaluating the multiaxial loading capacities of lattice structures and the impact of mesoscale design parameters, especially under unknown cyclic conditions during operation. To address this, we propose a data-driven methodology that integrates a multiscale shakedown evaluation method with a hybrid machine learning (HML) model. Our HML model, incorporating ensemble learning techniques and hyperparameter tuning via random search, accurately predicts the multiaxial shakedown fatigue loading capacity of a representative peanut-shaped auxetic lattice structure with parameterized geometry. The HML model's exceptional performance, demonstrated by a Normalized Root Mean Squared Error (NRMSE) of 0.018 and a coefficient of determination (R²) of 0.945, underscores its reliability, precision, and practicality. Additionally, sensitivity-based parametric analyses reveal the significant influence of center distance and edge width on the multiaxial fatigue properties of the lattice structure. This work offers an efficient tool for quantifying the contributions of various design parameters and loading conditions to multiaxial shakedown loading capacities.