Abastact. Pest recognition is essential to precision agriculture, enabling higher crop  yields and sustainable pest management. This issue has recently been addressed using deep learning–based approaches, but existing datasets suffer from insufficient species diversity, limited samples, and a lack of detailed taxonomic levels required for effective training. To address these limitations, we introduce HIVE, a large-scale benchmark dataset comprising 13,434 categories and 2,106,735 insect pest images annotated across 19 taxonomic levels. HIVE strengthens pest recognition by providing fine-grained morphological cues across diverse species and enables accurate classification across multiple hierarchical taxonomic levels. Moreover, HIVE achieves superior downstream performance on existing pest datasets, consistently surpassing models pretrained on widely used benchmarks. It also allows the retrieval of morphologically similar species from unseen pest images, providing practical support for non-expert users without specialized taxonomic knowledge. 

Abastact. Multi-modal approaches used for 3D CAD generation require substantial computational resources, necessitating efficient training. To address this, we propose GuideCAD, which leverages semantically rich visual-textual representations having only a small number of trainable parameters to generate 3D CAD models. Specifically, GuideCAD uses a mapping network that converts image embeddings into prefix embeddings, enabling a pretrained large language model to integrate visual and textual information. As a result, a transformer-based decoder predicts the construction sequence using the visual-textual embeddings in order to generate the 3D CAD model. For experimental evaluation, we construct a new dataset, referred to as GuideCAD, which consists of text-image pairs. Each pair includes a text prompt that represents a 3D CAD construction sequence and its corresponding 3D CAD image. Our experimental results show that GuideCAD generates comparably high-quality 3D CAD models while using approximately four times fewer parameters and achieving twice the training efficiency compared to fine-tuning approaches. We have released the source code and dataset for our method at: https://github.com/mskimS2/GuideCAD.

Abastact. The success of Transformer-based models has encouraged many researchers to learn CAD models using sequence-based approaches. However, learning CAD models is still a challenge, because they can be represented as complex shapes with long construction sequences. Furthermore, the same CAD model can be expressed using different CAD construction sequences. We propose a novel contrastive learning-based approach, named ContrastCAD, that effectively captures semantic information within the construction sequences of the CAD model. ContrastCAD generates augmented views using dropout techniques without altering the shape of the CAD model. We also propose a new CAD data augmentation method, called a Random Replace and Extrude (RRE) method, to enhance the learning performance of the model when training an imbalanced training CAD dataset. Experimental results show that the proposed RRE augmentation method significantly enhances the learning performance of Transformer-based autoencoders, even for complex CAD models having very long construction sequences. The proposed ContrastCAD model is shown to be robust to permutation changes of construction sequences and performs better representation learning by generating representation spaces where similar CAD models are more closely clustered. 

Abastact. In this study, we propose a new approach for automatically generating high-quality non-uniform meshes based on supervised learning. The proposed framework, GMR-Net, is based on training a graph convolutional neural network (GCN) that learns local error density around each interior vertex in the mesh. The target edge length is then estimated using the predicted local error density. GMR-Net performs vertex-based error estimation as a preprocessing step to avoid expensive error estimation. The proposed GCN-based neural net has new input features for learning the local error density with various polygonal geometries, a range of partial differential equation parameters, and boundary conditions. During the test process, it predicts the target edge length around each interior vertex for the coarse input mesh, and the final non-uniform meshes are generated over the domain using Gmsh. To demonstrate the effectiveness of the GMR-Net, we tested the GMR-Net on Poisson’s equations and linear elasticity problems to previously unseen and new polygonal geometries and a range of PDE coefficients and boundary conditions. 

Abastact. The existing mesh refinement approach generates an optimal finite element mesh in an iterative manner. However, this approach requires the user to have prior knowledge of complex PDE problems and error estimation methods. Furthermore, an expensive error estimation process has to be repeatedly performed. The present study was motivated by MeshingNet, a pioneering work that performs mesh refinement using neural nets. In MeshingNet, the proposed neural net learns and predicts a target area upper bound, which is the criterion to refine the elements, for a given PDE problem and geometry, and a high-quality non-uniform mesh is generated using this criterion. However, the neural net proposed in MeshingNet has limitations because it assumes that polygons with the same number of boundary vertices are used as inputs as well as outputs. In addition, the PDE error prediction method used is inefficient and lacks robustness. We propose a new mesh refinement framework for solving PDEs based on deep learning. The proposed framework models iterative mesh refinement problems as multi-class classification problems and proposes a new neural net for solving them. Unlike the previous studies, the proposed neural net is flexible in that it can be applied to geometry and boundary conditions having various polygonal shapes. We also propose efficient and robust methods that predict PDE errors using structured grids. The experimental results show that the proposed framework successfully generates high-quality non-uniform meshes, and the solution accuracy of these meshes is considerably improved compared to a uniform mesh with a similar number of elements. 

Abastact. In this paper, we propose a novel deep learning model that generates 3D CAD models represented as command sequences. The proposed model consists of a transformer-based autoencoder and a diffusion-based generative model. The transformer-based autoencoder learns latent representations to generate CAD command sequences effectively, and the diffusion-based generative model learns to generate latent variables for creating new command sequences. Experimental results show that the proposed model is able to successfully generate valid and new 3D CAD models. Our experiments also show that the proposed model learns the meaningful latent space for CAD models. 

Abastact. We propose a new mesh refinement framework that performs mesh refinement in a data-driven way for given PDEs, geometric domains, and boundary conditions. The proposed framework is able to produce non-uniform meshes from an initial uniform coarse mesh by reducing the element size at the location where the error is expected to be large. The proposed framework is able to learn and accurately predict the target element size without repeatedly performing the expensive post error estimation process. In the experiment, we test the proposed framework on the Poisson’s equation, which is a representative elliptical PDE. Our experimental results show that the proposed framework successfully produces non-uniform meshes and these non-uniform meshes have better efficiency and accuracy compared with the uniform meshes that have more number of elements.