A computational framework that transforms single images into 3D physical objects by considering mechanical properties, external forces, and rest-shape geometry. Unlike existing single-view 3D reconstruction methods that overlook these factors, our approach ensures physical compatibility by embedding these attributes into the reconstruction process and results in 3D shapes that exhibit desired physical behaviors and withstand real-world forces. Our framework is particularly useful for dynamic simulations and 3D printing, where physical compatibility is crucial.

A Lagrangian representation for reconstructing 3D shapes with high-quality geometry. Unlike conventional methods that rely on Eulerian representations, TetSphere splatting uses tetrahedral meshes to achieve superior mesh quality without neural networks or post-processing. This method deforms multiple initial tetrahedral spheres through differentiable rendering and geometric energy optimization, ensuring computational efficiency. TetSphere splatting excels in applications like single-view 3D reconstruction and image-/text-to-3D content generation, outperforming existing methods in optimization speed, mesh quality, and preservation of thin structures.

A novel skeletal representation that tackles the prevailing issues around compactness and reconstruction accuracy in existing skeletal representations. Our approach augments the continuous elements in the medial axis representation to effectively shift the complexity away from discrete elements. We introduce generalized enveloping primitives, an enhancement of the standard primitives in medial axis, which ensures efficient coverage of intricate local features of the input shape and substantially reduces the number of discrete elements required. Moreover, we present a computational framework that constructs a medial skeletal diagram from an arbitrary closed manifold mesh. We exemplify the versatility of our representation in downstream applications such as shape optimization, shape generation, mesh decomposition, mesh alignment, mesh compression, and user-interactive design.

A graph grammar-based model to efficiently represent and reason over these molecules, leveraging hierarchical design spaces and motifs. By utilizing random walks over the design space, our method excels in molecule generation and property prediction. We demonstrate superior performance, efficiency, and synthesizability of predicted molecules compared to existing methods, while offering detailed chemical interpretability.

A bilevel optimization framework that combines LLMs’ reasoning abilities with the computational power of simulations. In this framework, LLMs propose hypotheses and reason about discrete components, while simulations provide feedback and optimize continuous parameters. Our experiments demonstrate the efficacy of SGA in discovering constitutive laws and designing molecules, revealing novel and coherent solutions beyond conventional human expectations.

A novel method for computing smooth, inversion-free maps based on variational quasi-harmonic maps and a special Cauchy boundary condition. By transforming the mapping problem into an optimal control problem defined by an elliptic PDE and minimization within a family of functionals, an efficient numerical procedure is proposed, yielding significant improvements in speed and quality over current methods and allowing for optimization of a generic energy in the framework of quasi-harmonic maps.

A data-efficient property predictor for molecular properties utilizing a learnable hierarchical molecular grammar and graph neural diffusion over the grammar-induced geometry. Despite limited labeled data, the proposed method significantly outperforms various baselines, including supervised and pre-trained graph neural networks, in both small and large datasets.

A data-efficient generative model that can be learned from datasets with orders of magnitude smaller sizes than common benchmarks. At the heart of this method is a learnable graph grammar that generates molecules from a sequence of production rules. Our learned graph grammar yields state-of-the-art results on generating high-quality molecules for three monomer datasets that contain only ∼20 samples each.

A formal grammar specifically designed for the polymers, providing a concise, explicit, and robust representation for any polymer within a given class. This project serves as a foundational blueprint for constructing general chemical design models, with closely integrated representation and generative capabilities.

A framework for variable-context multi-objective optimization capturing Pareto fronts over a range of contexts (Pareto gamut). A global-local optimization algorithm is developed to discover the Pareto gamut. Its practical utility is demonstrated for several engineering design problems.

Take the first step to understand adversarial robustness of neural networks from an architectural perspective. Based on our “robust architecture Odyssey”, we design a family of robust architectures RobNets that exhibit superior robustness performance to other widely used architectures.

An economical solution that learns the data augmentation policy along with network training. The augmentation policy is formulated as a parameterized probability distribution, thus allowing its parameters to be optimized jointly with network parameters. Compared with SOTA methods, this framework achieves 60× faster on CIFAR-10 and 24× faster on ImageNet, while maintaining competitive accuracies.

An inverse reinforcement learning framework for the training of an agent to learn to search network structures that are topologically inspired by human-designed network. A mirror stimuli function motivated by biological cognition theory is proposed to extract the Summary topological knowledge and facilitates the discovery of architectures that stride a satisfactory balance between accuracy and efficiency.

A deformable face tracking system that can automatically balance the efficiency and effectiveness according to the image quality and tracking complexity of each frame. Dual-agent deep reinforcement learning is used to make adaptive decisions to position updates of bounding boxes and landmarks. The interaction of these two subtasks is explicitly exploited following a Bayesian model.