Date of Award:


Document Type:


Degree Name:

Master of Science (MS)


Mechanical and Aerospace Engineering

Committee Chair(s)

Hailei Wang


Hailei Wang


Geordie Richards


Barton Smith


Jia Zhao


In order to reduce the cost ($) per megawatts hour (MWh) of electrical energy generated by a nuclear power cycle with a novel small modular reactor (SMR), a new SMR-based nuclear power cycle with Methanol as working fluid was designed. It was built virtually with the Python & Coolprop software based on all components’ physical properties, and it is therefore called the physics-based model. The physics-based model would require seven user-defined values as input for the seven free design parameters, respectively. The physics-based model outcomes include LCOE (the cost per megawatts hour of electrical energy generated by the cycle), the first-law efficiency of the power cycle (the ratio between the net power out of the power cycle and the net thermal energy into the power cycle), and the penalty (severity of system violation with the given design parameters values). In order to compare between different power cycles, the corresponding optimized LCOE are needed. However, in order to find the design parameters that optimize the system LCOE, it can take up to three days with the physics-based model, because the physics-based model is highly complex and it takes thousands of iterations averagely to the optimize the power system.

In confronting the time complexity issue in optimization, the study in this paper explores the viability of replacing the physics-based model with a machine learning-based surrogate model. During the optimization procedure, the machine learning-based surrogate model is expected to accelerate process of finding the corresponding outcomes, and thus to save time. Candidate surrogate models are built and analyzed in terms of their prediction accuracy. The last chosen model is further optimized in terms of its structure and hyper-parameters. With the structurally and parametric optimized surrogate model being incorporated, different global optimizers are used and analyzed. As the result, the optimized design parameters from the surrogate-optimizer model are fed into the physics-based model, and their corresponding results are compared with the baseline optimized results of the physics-based model.

In conclusion, the study reveals that the Multilayer Perceptron (MLP) networks with two hidden layers gives the best prediction performance, and therefore they are chosen as the surrogate model. In addition, four global optimizers, namely the basinhopping, the differential evolution, the dual annealing and the fmin, are working well along with the chosen surrogate model. They integrated surrogate-optimizer model is capable of finding the optimized LCOE as well as the corresponding design parameters. In comparison with the baseline optimized LCOE, the relative error is less than 3.5%, and this searching procedure completed within 30 minutes.