A Non-deterministic propeller design optimization framework leveraging machine learning based boundary element methods surrogates

Marine propeller design is a cornerstone of naval architecture and marine engineering, serving as a classic optimization problem that has captivated the attention of scholars, engineers, and practitioners for generations. The challenge lies in developing efficient, reliable, and cost-effective propellers that excel in a variety of operating conditions while adhering to strict environmental regulations and addressing noise and vibration concerns. As a critical component in the propulsion systems of ships, submarines, and other marine vessels, the marine propeller’s design directly impacts vessel performance, fuel consumption, and emissions. Marine propellers often operate under uncertain conditions, including inflow, rate of revolutions, and manufacturing tolerances. On the one hand, a deterministic design approach that does not consider these stochastic inputs can lead to excessive sensitivity to slight variations in operational conditions, resulting in a geometry that is sub-optimal for real-world functioning scenarios. On the other hand, the computational effort required to quantify the design’s uncertainties may be prohibitive even when mid-fidelity solvers, like Boundary Element Methods (BEM), are employed in Simulation-Based Design Optimization. As in the case of deterministic design optimizations driven by mid-fidelity codes, Machine Learning methodologies represent a computational booster of the procedure. By realizing computationally cheap yet accurate surrogate models of the key performance indicators of the design, they allow for the hundreds of thousands of calculations needed by sampling methods to evaluate the uncertainty of the design and drive the process towards configurations less sensitive to inputs variations, making the non-deterministic design optimization a feasible alternative to conventional deterministic design-by-optimization methodologies. In this paper, deterministic and non-deterministic designs are carried out in the case of a conventional propeller, considering uncertainties of the nominal functioning conditions. BEMs computations are used to train accurate Machine Learning-based surrogate models. Thanks to their cost-effectiveness in the forward phase, they enable the assessment of the uncertainties necessary for a non-deterministic design optimization framework. In order to demonstrate the efficacy of the proposed methodology, the optimal geometries attained through the utilization of surrogates are re-tested utilizing medium-fidelity BEM solvers to validate the superior performance of the non-deterministic design approach.