Although deep learning-based point cloud denoising methods have demonstrated impressive results in recent years, their effectiveness usually degrades substantially in real-world scenarios where the noise distribution of test point clouds deviates from that of the training set. To address this challenge, we propose a test-time adaptive denoising (TTAD) framework that enables existing denoising approaches to adapt to unknown noise patterns at test time without relying on additional supervision or retraining from source data. Taking a pre-trained denoising model and a test noisy point cloud as input, TTAD first employs a residual-driven noise modeling (RDNM) module to estimate the sample-specific noise distribution through analyzing the deviation between the noisy input and its pseudo-clean version denoised by the initial model. Then, a noise-aware contrastive adaptation (NACA) module is designed to build sample-specific noisy-noisier test-time training pair from the denoised point cloud, and update the model parameters over a few adaptation steps per sample at test time through contrastive learning. Furthermore, TTAD naturally extends to a continual TTAD (CTTAD) setting, facilitating efficient progressive refinement across sequential test samples with consistent noise distributions, while reducing the computational cost per sample. Extensive experiments on both synthetic and real-world datasets demonstrate that our approach significantly improves the generalization ability of existing denoising models under unseen noise, achieving superior performance across multiple evaluation metrics.
Triply periodic minimal surface (TPMS)-based structures are increasingly being developed into multi-topology heterogeneous designs to meet multifunctional requirements. However, geometric mismatches at heterogeneous interfaces often induce stress concentration and structural failure, making interfacial connectivity a critical design bottleneck. To address this challenge, this study proposes a data-driven multiscale continuous topology optimization method for heterogeneous TPMS structures. First, a heterogeneous unit-cell library comprising five typical TPMS configurations is established. A weighting function-based fusion strategy is then introduced to precisely model the interface transition zones, thereby achieving smooth geometric transitions among multiple configurations. Second, to capture the complex physical nonlinearities within the transition zones, a multilayer perceptron (MLP) surrogate model is developed to establish a high-fidelity mapping between the design parameters and the equivalent elastic tensors. Building upon this, a multiscale concurrent topology optimization framework incorporating geometric connectivity constraints is proposed, which simultaneously balances macroscopic mechanical performance and microscopic geometric continuity. Numerical examples are presented to verify the effectiveness and scalability of the proposed method. Parametric analyses indicate that the smooth transition strategy effectively alleviates interfacial stress concentrations. Furthermore, the multiscale topology optimization results demonstrate that the heterogeneous designs successfully break the mechanical performance limits of single-topology configurations under identical volume constraints, while maintaining strict microstructural geometric continuity. This study provides a solid theoretical framework for achieving smooth interfacial connectivity and high-performance tailored design of heterogeneous metamaterials.
Porous structures with tunable mechanical properties are widely required across personalized biomedical devices, orthotic applications, and protective cushioning systems, yet conventional design methods relying on empirical iteration and finite element analysis (FEA) are time-consuming and inefficient for rapid, patient-specific customization. To address this limitation, this study proposes a deep learning–based inverse design framework for Gyroid triply periodic minimal surface (TPMS) porous structures, enabling direct identification of material type and porosity from target stress–strain responses. The framework employs a hierarchical neural network comprising stress-level classification, inverse parameter identification, and forward response reconstruction modules, with stress–strain curves represented by 200 paired data points. Physics-constrained multi-objective loss functions are incorporated to enforce monotonicity, smoothness, and densification behavior, ensuring physically consistent predictions. The model is trained using uniaxial compression datasets generated by finite element simulations, with material parameters calibrated to reported experimental data. Five biocompatible polymeric materials (PA12, PA11, TPE, TPU95A, and TPUA09) and porosities ranging from 50% to 80% are considered. Validation results show a material classification accuracy of 94.8% and a mean porosity prediction error of 3.2%. The reconstructed stress–strain curves closely match target responses across elastic, plateau, and densification regimes, achieving a composite quality score of 94.9/100. Finite element cross-validation further confirms the reliability of the predicted design parameters. This work demonstrates a rapid and data-driven approach for inverse design of mechanically tunable porous structures, with broad applicability to personalized biomedical devices, protective equipment liners, and other engineering applications requiring spatially customized mechanical responses.
The design of the microstructure of porous materials has important scientific significance and a wide range of practical application prospects in determining their macroscopic physical properties, which makes modelling stochastic microstructure of porous materials significant for the investigation and innovative use of materials in the scientific and engineering fields. With the advancement of deep learning, the generative adversarial network (GAN) has become widely utilized as a powerful generative model. It provides technical support for the stochastic microstructure reconstruction of porous materials. However, existing GAN suffers from lack of conditional data guidance, gradient vanishing, and gradient explosion. To address these issues, we introduce a new method that utilizes conditional sampling points of the microstructure of porous materials as conditional data to guide the GAN in generating samples with high similarity, thereby enhancing the quality of model reconstruction. Additionally, we also introduce a gradient normalization (GN) module to stabilize the model's training process, effectively addressing the issues of gradient vanishing and exploding. We use this new method and other reconstruction methods to reconstruct porous materials with stochastic microstructure, and compare them in various evaluation metrics such as morphological measurements, multi-point connectivity, and Fréchet inception distance. Furthermore, the scalability of GN-CGAN is demonstrated by generating a 192 × 192 × 192 Berea sandstone microstructure from a 128 × 128 × 128 exemplar, highlighting its potential for large-volume stochastic reconstruction. The results show that our reconstruction method performs well across multiple evaluations.
An efficient point-in-body query method on closed triangulated surfaces is presented. It is targeted for the initialization of volume fractions within multi-material simulations. It uses the winding number of the surface to determine whether a point is inside its volume or not. Its key point is the hierarchical evaluation of the solid angle contributions, using a tree constructed from a recursive planar subdivision of the domain. Its accuracy is comparable to a cad kernel, while faster on complex shapes.
