Search for: All records

Abstract The average lifetime or the mean time to failure (MTTF) of a product is an important metric to measure the product reliability. Current methods of evaluating the MTTF are mainly based on statistics or data. They need lifetime testing on a number of products to get the lifetime samples, which are then used to estimate the MTTF. The lifetime testing, however, is expensive in terms of both time and cost. The efficiency is also low because it cannot be effectively incorporated in the early design stage where many physics-based models are available. We propose to predict the MTTF in the design stage by means of a physics-based Gaussian process (GP) method. Since the physics-based models are usually computationally demanding, we face a problem with both big data (on the model input side) and small data (on the model output side). The proposed adaptive supervised training method with the Gaussian process regression can quickly predict the MTTF with a reduced number of physical model calls. The proposed method can enable continually improved design by changing design variables until reliability measures, including the MTTF, are satisfied. The effectiveness of the method is demonstrated by three examples. 

Abstract This paper presents an adaptive Kriging based method to perform uncertainty quantification (UQ) of the photoelectron sheath and dust levitation on the lunar surface. The objective of this study is to identify the upper and lower bounds of the electric potential and that of dust levitation height, given the intervals of model parameters in the one-dimensional (1D) photoelectron sheath model. To improve the calculation efficiency, we employ the widely used adaptive Kriging method (AKM). A task-oriented learning function and a stopping criterion are developed to train the Kriging model and customize the AKM. Experiment analysis shows that the proposed AKM is both accurate and efficient. 

Abstract Average lifetime, or mean time to failure (MTTF), of a product is an important metric to measure the product reliability. Current methods of evaluating MTTF are mainly statistics or data based. They need lifetime testing on a number of products to get the lifetime samples, which are then used to estimate MTTF. The lifetime testing, however, is expensive in terms of both time and cost. The efficiency is also low because it cannot be effectively incorporated in the early design stage where many physics-based models are available. We propose to predict MTTF in the design stage by means of physics-based models. The advantage is that the design can be continually improved by changing design variables until reliability measures, including MTTF, are satisfied. Since the physics-based models are usually computationally demanding, we face a problem with both big data (on the model input side) and small data (on the model output side). We develop an adaptive supervised training method based on Gaussian process regression, and the method can then quickly predict MTTF with minimized number of calling the physics-based models. The effectiveness of the method is demonstrated by two examples. 

This paper presents a modeling and uncertainty quantification (UQ) study of the photoelectron sheath near the lunar surface. A fully kinetic 3-D finite-difference (FD) particle-in-cell (PIC) code is utilized to simulate the plasma interaction near the lunar surface and the resulting photoelectron sheath. For the uncertainty quantification analysis, this FD-PIC code is treated as a black box providing high-fidelity quantities of interest, which are also used to construct efficient reduced-order models to perform UQ analysis. 1-D configuration is chosen to present the analytic sheath solution as well as to demonstrate the procedure and capability of the UQ analysis.