Bayesian Optimization of Genetic Algorithm Hyperparameters in a Multi-Fidelity Framework for Efficient Lattice Material Design

Designing high-performance lattice materials often requires genetic algorithms (GAs) to navigate complex, multi-modal design spaces. However, the performance of these GAs is highly sensitive to hyperparameters like population size, mutation rates, and selection pressure. Tuning these parameters is typically a computationally expensive, black-box process. This study asks whether a multi-fidelity Bayesian optimization (BO) framework can automate this tuning process to improve efficiency and mechanical performance in lattice design.

The authors developed a three-level hierarchical framework to optimize GA hyperparameters. At the lowest fidelity, a Gaussian process (GP) surrogate within a BO framework guides the search for optimal hyperparameters. At the medium fidelity, a 3D convolutional neural network (3DCNN) surrogate accelerates the evaluation of candidate lattice architectures. Finally, at the high-fidelity level, FFT-based computational homogenization validates the mechanical properties of the optimized structures. The study systematically evaluated different BO acquisition functions, identifying logNEI as the most effective for handling the inherent noise in GA evaluations.

The proposed framework successfully identified hyperparameter configurations that allow a 25-generation GA run to achieve elastic modulus values comparable to a standard 75-generation run. By introducing a penalized BO objective, the researchers balanced mechanical performance against the number of required lattice evaluations, enabling a 24% reduction in total computational cost (from 225 to 171 hours). The optimized hyperparameters also eliminated the need for lattice mutation, simplifying the evolutionary process without sacrificing structural performance.