Improving active learning in multi-fidelity hydrodynamic optimization

Despite recent advances in machine learning, simulation-driven design optimization using high-fidelity simulations may still be prohibitively expensive for practical applications. This paper investigates improvements in multi-fidelity surrogate-based hydrodynamic optimization, which are intended to make the process faster and more efficient. Specific innovations are: a) the use of a reduced initial dataset with only one data point for all fidelity levels except the lowest, to reduce the computational cost of surrogate-model initialization; b) accounting for noise variance in the selection of the fidelity level to sample, to avoid oversampling well-resolved but noisy fidelity levels, c) improving the automatic mesh adaptation protocol for the CFD simulations, to further optimize mesh adaptation; and d) restarting highfidelity simulations from converged low-fidelity results, to improve the overall efficiency of the design optimization process. These methodological advancements are demonstrated for an analytical test problem, as well as the shape optimization of a NACA 4-digit airfoil and the DTMB 5415 for calm-water resistance. The results show that the reduced dataset drastically reduces the computational cost of initialization and favors efficient low-fidelity exploration of the design space. The noise-corrected fidelity selection encourages the selection of higher fidelities, to effectively determine the true optimum. Finally, the CFD solver advancements make high-fidelity simulations faster, up to eight times. Compared with previous work, the solution of the DTMB 5415 problem exhibits a more robust training process, providing a slightly improved optimum, in less than half the computational time.