Parametric simulation optimization and data space refinment using LHS and surrogate machine learning models

Parametric simulations in engineering are critical for exploring the effects of various parameters on the performance of a system. As the number of parameters and variations increases, the computational burden of running full-scale parametric simulations becomes prohibitively expensive. This thesis investigates the efficacy of training machine learning models on results of parametric simulations using Latin Hypercube Sampling (LHS) which enables efficient exploration of the data space with fewer simulations. By training surrogate models on these reduced datasets, it is possible to approximate the behavior of the system without needing to perform every possible simulation. Then, the surrogate models can be used to quickly refine the parameter space, by minimizing an objective function of interest. Once the minimum is located, a more granular parametric analysis can be performed, focusing on the relevant regions of the design space and refined distribution of each parameter, allowing for a significant reduction in computational cost while maintaining accuracy. The study evaluates the effectiveness of this approach in engineering contexts and discusses the trade-offs between computational efficiency and result fidelity.