Single and Multiresponse Adaptive Design of Experiments with Application to Design Optimization of Novel Heat Exchangers

Abstract

Engineering design optimization often involves complex computer simulations.

Optimization with such simulation models can be time consuming and sometimes

computationally intractable. In order to reduce the computational burden, the use of

approximation-assisted optimization is proposed in the literature. Approximation

involves two phases, first is the Design of Experiments (DOE) phase, in which

sample points in the input space are chosen. These sample points are then used in a

second phase to develop a simplified model termed as a metamodel, which is

computationally efficient and can reasonably represent the behavior of the simulation

response. The DOE phase is very crucial to the success of approximation assisted

optimization.

This dissertation proposes a new adaptive method for single and multiresponse

DOE for approximation along with an approximation-based framework for multilevel

performance evaluation and design optimization of air-cooled heat exchangers.

The dissertation is divided into three research thrusts. The first thrust presents a new

adaptive DOE method for single response deterministic computer simulations, also

called SFCVT. For SFCVT, the problem of adaptive DOE is posed as a bi-objective

optimization problem. The two objectives in this problem, i.e., a cross validation error

criterion and a space-filling criterion, are chosen based on the notion that the DOE

method has to make a tradeoff between allocating new sample points in regions that

are multi-modal and have sensitive response versus allocating sample points in

regions that are sparsely sampled. In the second research thrust, a new approach for

multiresponse adaptive DOE is developed (i.e., MSFCVT). Here the approach from

the first thrust is extended with the notion that the tradeoff should also consider all

responses. SFCVT is compared with three other methods from the literature (i.e.,

maximum entropy design, maximin scaled distance, and accumulative error). It was

found that the SFCVT method leads to better performing metamodels for majority of

the test problems. The MSFCVT method is also compared with two adaptive DOE

methods from the literature and is shown to yield better metamodels, resulting in

fewer function calls.

In the third research thrust, an approximation-based framework is developed for

the performance evaluation and design optimization of novel heat exchangers. There

are two parts to this research thrust. First, is a new multi-level performance evaluation

method for air-cooled heat exchangers in which conventional 3D Computational

Fluid Dynamics (CFD) simulation is replaced with a 2D CFD simulation coupled

with an e-NTU based heat exchanger model. In the second part, the methods

developed in research thrusts 1 and 2 are used for design optimization of heat

exchangers. The optimal solutions from the methods in this thrust have 44% less

volume and utilize 61% less material when compared to the current state of the art

microchannel heat exchangers. Compared to 3D CFD, the overall computational

savings is greater than 95%.