residual deformation learning

Addressing the challenges of scarce annotated data and severe simulation-to-reality domain shift in structural deformation detection under real-world conditions, this paper proposes a controllable physics-based simulation framework integrated with lightweight domain adaptation. Methodologically: (1) we design a physics-inspired, controllable deformation image generator tailored for deformation classification, enabling fine-grained adjustment of deformation parameters; (2) we introduce a feature-level adversarial domain adaptation network to align simulated feature distributions with those of real-world data; (3) the model achieves efficient fine-tuning using only a small number of real deformation samples. Experiments demonstrate that our approach improves classification accuracy by 12.6% over a pure-simulation baseline on cross-domain deformation classification tasks, significantly alleviating data dependency. The source code is publicly available.

Traditional linear reduced-order models (ROMs) rely on fixed subspaces, rendering them inadequate for capturing nonlinear geometric deformations under large displacements, thereby limiting accuracy. This paper proposes an adaptive linear ROM framework that dynamically updates the reduced-order mapping via an online error-driven adaptation mechanism. By leveraging Grassmann manifold interpolation, the method enables robust interpolation and update of the reduced basis over the subspace manifold spanned by historical displacement snapshots. This approach explicitly relaxes the static subspace assumption inherent in conventional ROMs, achieving a favorable balance between computational efficiency and approximation fidelity. Experimental results demonstrate that, at comparable computational cost, the proposed method significantly reduces simulation error relative to PCA-based linear ROMs, markedly improving both accuracy and numerical stability in large-deformation scenarios.

Traditional spatial deformation methods struggle to model covariate-driven nonstationary spatial dependencies and exhibit limited generalization. This work proposes a covariate-driven diffeomorphic spatial deformation framework that represents the deformation as a function of covariates, generating smooth and invertible mappings via velocity fields in a Lie algebra. To enhance stability and generalizability, the method incorporates a physics-informed truncation strategy for high-order interaction terms. It is the first approach to enable nonstationary Gaussian process extrapolation under covariate conditioning, demonstrating superior small-sample predictive performance on both synthetic data and real-world applications in manufacturing and geostatistics.

To address the accuracy-efficiency trade-off in r-adaptive mesh relocation within the finite element method (FEM), this paper proposes an end-to-end, error-driven optimization framework based on graph neural networks (GNNs). We introduce the first GNN-integrated online r-adaptivity pipeline for FEM, wherein a learnable policy directly minimizes the numerical error of partial differential equation (PDE) solutions—replacing conventional paradigms reliant on nonlinear PDE solves and hand-crafted error estimators. A geometry-aware graph representation ensures alignment between mesh and solution spaces, and the framework is implemented and trained end-to-end within the Firedrake environment. Experiments demonstrate that our method achieves several-fold computational speedup over traditional approaches while significantly reducing numerical error, outperforming both classical r-adaptive schemes and state-of-the-art machine learning–based adaptive mesh methods in overall accuracy-efficiency balance.

为解决物理量恢复难题，提出RealSimLoop框架，通过视觉反馈和降阶神经子空间内的可微仿真，实现在线实到仿真的高效自适应。

该研究提出了一种统一的基于样本的高斯编码方法，用于处理结构化、非结构化及粒子数据，通过优化高斯基元减少数据量并提高重建精度。

该研究通过结合RVQ、U-Net像素空间残差校正和GAE误差界限保证的方法，解决了科学数据压缩中重建误差控制的问题。

This work addresses the limitations of conventional physics-based simulation models, which typically assume material homogeneity and isotropy, thereby failing to capture subtle anisotropy and heterogeneity present in real-world objects and hindering simulation accuracy and sim-to-real transfer. To overcome this, the authors propose the MoSA framework, which builds upon a calibrated isotropic model and introduces a physically meaningful residual stress operator to represent weak material heterogeneities. MoSA innovatively integrates physical priors with data-driven learning through a microfacet-constrained cascaded neural network, while incorporating spatiotemporal derivatives of the deformation field as motion constraints to enhance dynamic consistency. Experimental results demonstrate that MoSA significantly improves modeling accuracy, generalization, and robustness, enabling more reliable sim-to-real transfer in robotic manipulation tasks.

This work addresses the challenges of physical inconsistency and poor generalization in dynamic prediction for deformable objects by proposing a material-aware, physics-corrected residual world model. The approach integrates a differentiable Material Point Method (MPM) physics simulator with two lightweight feedforward networks: Material from Motion (MfM) infers particle-level elastic parameters from visual inputs to enable online material identification for novel objects, while Residual from Dynamics (RfD) learns and corrects systematic simulation biases. Coupled with an uncertainty-guided active exploration mechanism, the model maintains physical consistency while achieving strong generalization. Experiments demonstrate that the method outperforms state-of-the-art approaches in prediction accuracy on real-world deformable object manipulation sequences, with well-calibrated confidence estimates that effectively support downstream decision-making.

This work addresses the challenge in computational solid mechanics of simultaneously preserving irreversible plastic deformation and continuous damage evolution during complex loading–unloading–reloading cycles. To this end, the authors propose a compact, vectorizable explicit update algorithm that couples elastoplasticity and damage. The method employs a softplus-based equivalent plastic strain combined with a maximum historical projection to enforce irreversibility, and introduces an exponential scalar degradation variable to track damage evolution. Within a unified energy framework, both active and frozen solution branches are resolved analytically, thereby eliminating the need for local Newton iterations. Numerical experiments demonstrate high accuracy under proportional or near-proportional loading (with only 1.53% error), implementation simplicity, and gradient consistency. Although error increases under large path-angle reversals, the approach achieves enhanced smoothness and algorithmic robustness at minimal additional computational cost.