Dynamic performance of energy-absorbing box with auxetic cellular structures using parametric modeling and machine learning

To overcome the poor transferability of conventional optimization results for auxetic cellular structures (ACS) energy-absorbing boxes, a dimension-adaptive design framework integrating parametric modeling and machine learning is proposed herein. An energy-absorbing box filled with a three-dimensional lower butterfly wing cellular structure is adopted as the study object. Firstly, a parametric simulation platform is established in Abaqus and validated by drop-hammer impact tests. Subsequently, 600 training samples are generated with the adoption of Latin hypercube sampling with a fluctuation-enhancement strategy, and a neural network surrogate model (NNSM) is developed to rapidly predict Specific Energy Absorption (SEA) and Peak Crushing Force (PCF). Furthermore, efficient multi-objective optimization can be realized by further combining the NNSM with the non-dominated sorting genetic algorithm II. In addition, the synergistic mechanism between the outer shell and the internal ACS is revealed. Based on the NNSM, the effects of design parameters on structural performance and the constraining role of overall dimensions on performance boundaries are further explored. Results from four optimization cases demonstrate that the average prediction errors of the NNSM for SEA and PCF are merely 2.73 % and 1.46 %, respectively. Compared with the initial designs, an average increment of 8.66 % in SEA and an average decrease of 1.45 % in PCF are achieved for the optimized structures. The effectiveness of the proposed dimension-adaptive design framework is validated by the obtained results, and a novel insight can be provided for the optimal design of auxetic energy-absorbing structures.