The expected hypervolume improvement (EHVI) is one of the most popular infill criteria for multiobjective optimization problems. Although it has a significant advantage in exploring potential Pareto-optimal solutions, it has rarely been applied in many-objective problems due to its high computational cost. To address this issue, this article proposes an expensive many-objective optimization algorithm based on the framework of nondominated sorting genetic algorithm III (NSGA-III) and assisted by the kriging surrogate models. In the proposed algorithm, the Monte Carlo sampling (MCS) method for EHVI estimation is improved by importance sampling, in which only one sampling process is required during the entire optimization process using a uniform distribution in normalized objective space. Considering the predicted uncertainty from the kriging model, an uncertainty-assisted nondominated sorting approach is proposed to substitute for the conventional approach in NSGA-III. In the proposed method, the predicted uncertainty is incorporated into the objective space as one independent dimension for nondominated sorting, which can enable the exploration of potential points with desirable EHVI values. In addition, the proposed algorithm considers the diversity of the solutions by de-emphasizing the pursuit of the best EHVI. The experimental results on benchmark problems demonstrate that the proposed EHVI calculation method can save computational costs compared with MCS and indicate the superiority of the proposed algorithm over the others.

NVIDIA Omniverse offers a unified platform for 3-D production pipelines based on the digital twins of real physical systems. The Internet of Things facilitates the acquisition of Omniverse data from different information sources, including finite-element model simulations and various sensors, and such data describe the responses of physical systems. According to their response description accuracies, these multisource data can be divided into different fidelity levels. High-fidelity (HF) data describe responses of the given system accurately but are costly to obtain. In contrast, low-fidelity (LF) data are inexpensive but often do not reach the desired accuracy level. Multifidelity data fusion (MDF) aims to use massive LF data and small amounts of HF data to develop the digital twin of a real physical system to produce accurate digital system responses. In this article, we propose a novel generative adversarial network for MDF. Experimental results show that the proposed model performs better than the state-of-the-art methods without any specific assumptions regarding the data distribution or data structure and has higher stability when addressing varying amounts of HF and LF data, especially in cases with very few HF data.