Multi-objective optimization for materials design with improved NSGA-II

Artificial intelligence and data science have accelerated the research and development into new materials. However, in a new materials design, it is more important to determine what the next possible “action” is (which is occasionally recommended to be one or more experimental points), than just the property predictions. This reason, multiple objective optimization, using a non-dominated sorting genetic algorithm (NSGA), for example, can be applied. However, a traditional NSGA mainly focuses on convergence and the diversity of the solutions, while neglecting the property preference expected by material scientists. Therefore, based on a Pareto optimal solution and NSGA-II algorithm, we propose an improved NSGA-II algorithm, called NSGA-II with preference (NSGA-IIP). At the main idea here, when the NSGA-IIP algorithm calculates the crowding distance, the cosine similarity is introduced to make the population of the next generation gather toward the direction of preference. For NSGA-IIP, it can adjust the solution set direction according to the expected property preferences. Moreover, we propose four indicators, i.e., the generational distance, inverted generational distance, mean cosine similarity, and points in sector to evaluate the optimization algorithm in terms of the convergence, solution diversity, and consistency regarding the preference. Finally, we apply this method to the materials design of a thermal barrier ceramic coating and compare it with the previous method. The results show that our proposed NSGA-IIP achieves better practicability and time performance for a new material design.