Recent advancements in metal additive manufacturing (AM) technologies have enabled the precise fabrication of complex designs, promoting integration between topology optimization and AM. However, AM processes like laser powder bed fusion (LPBF) inevitably lead to microstructural defects, including keyhole-induced porosity and lack of fusion porosity. These pores significantly degrade the mechanical strength of printed components. The complexity of the process leads to the generation of random and unpredictable pores, which can vary from part to part even if all process parameters and geometry remain the same. Thus, this study proposes a reliability-based topology optimization framework that incorporates the uncertainty arising from random porosity defects during the AM process in a computationally efficient way. In this study, the probability of pore formation is assumed to increase linearly with distance from the build plate, which serves as a proxy for temperature. Assuming that defects are induced exclusively by overheating, the likelihood of pore formation increases with the temperature rise during the build process. A Monte Carlo approach is used to account for the stochastic pore realizations. To avoid performing a separate FEA for each realization, analytical stress fields around pores are used to augment the stress field obtained from a single FEA per optimization iteration. Adjoint sensitivity analysis is then employed for efficient gradient evaluation. Through 2D numerical examples on regular Cartesian grids, this study demonstrates and evaluates the effectiveness of the proposed approach while accounting for process-dependent stochastic porosity.
The design of non-uniform curved grid-stiffened structures on complex surfaces presents a persistent trade-off: projection-based and structured-parameterization methods introduce geometric distortion or may fail under strongly curved geometries, while approaches capable of generating richly varied stiffener layouts, including non-uniform curved grid configurations, remain largely confined to planar or analytically simple geometries. This paper presents a unified framework that resolves this integration gap through three mutually enabling components. (1) Discrete conformal parameterization maps triangulated surfaces that are open and of zero genus onto a planar parametric domain, eliminating the need for analytical surface equations or CAD representations. (2) An explicit optimization method concurrently optimizes stiffener orientation, cross-sectional size, and spatial period within this domain, thereby extending non-uniform curved grid-stiffened texture generation to complex surfaces. (3) A reformulated anisotropic Helmholtz-based extrusion method treats the optimized 2D texture density field as a Dirichlet-type source term and extrudes it directly along the surface normal, in contrast to the conventional use of this equation as a density filter or extrusion constraint (Wang et al. 2020a), seamlessly bridging parametric design and 3D structural analysis without auxiliary constraints. The framework is validated on three examples of increasing geometric complexity: a sheet metal panel, a hyperbolic saddle surface, and a passenger aircraft nose. The proposed approach offers a systematic and computationally efficient solution for automating stiffener layout design on complex surfaces.
The next-generation aircraft aims to achieve high stealth performance while maintaining aerodynamic efficiency. To this end, high-fidelity numerical optimization is essential. The mean Radar Cross Section (RCS) is a common stealth metric in stealth optimization, but it neither directly reflects radar detection probability nor controls RCS fluctuation with aspect angles. This study replaces the mean RCS with the mean radar detection range at a given detection probability, thereby introducing a mechanism to adjust RCS fluctuation characteristics. To support this strategy, a radar detection probability model is established that accounts for single-pulse and multi-pulse coherent/non-coherent integration detection schemes, along with a non-parametric target fluctuation model that considers random angular distribution. The proposed methods are numerically validated through two aero-stealth optimization cases of a flying wing aircraft with an 18 m wingspan, which share identical aerodynamic strategies but differ in stealth metrics, with one using the mean RCS and the other using the mean detection range. The surrogate-based optimization algorithm with parallel infill-sampling strategy and adaptive design space expansion is employed for solutions within a large initial design space. The results indicate that the proposed strategy can distinguish designs with similar mean RCS but different fluctuation characteristics and achieve maximum mean detection range reductions of 6
Surrogate modeling is widely used to reduce the computational cost of high-fidelity simulations in complex engineering design. However, the predictive performance of individual surrogate models varies with problem dimensionality, nonlinearity, and sample distribution. Existing ensemble methods may also be affected by poorly performing component models, one-shot construction of the screened model library, and weighting strategies based mainly on either global or local information. To address these limitations, this study develops a screening-based sequential ensemble surrogate modeling method with dual-adaptive weighting, termed SEDA. Normalized prediction error sum of squares (PRESS) values are first used to remove clearly inferior models. The retained models are then ranked according to cross-validation performance and progressively incorporated to construct nested candidate libraries. A local uncertainty estimator based on the prediction dispersion of leave-one-out sub-models is introduced, and global weights derived from PRESS are combined with local weights determined from prediction-point-dependent uncertainty. The candidate ensembles are subsequently evaluated, and the model with the minimum internal prediction error is selected as the final predictor. SEDA is assessed using 26 benchmark functions and compared with five individual surrogate models and nine representative ensemble methods. Numerical results and statistical analyses show that SEDA achieves the best overall mean ranks and significantly outperforms most competing methods. Its engineering applicability is further examined through a simulation-based butterfly-valve case and an experiment-based safety-valve case. The results indicate that SEDA generally provides higher prediction accuracy and stronger robustness than the compared methods in the investigated engineering applications.
