Generative Deep Learning Models Research Articles

Inexpensive High Fidelity Melt Pool Models in Additive Manufacturing Using Generative Deep Diffusion

Defects in laser powder bed fusion (L-PBF) parts often result from the meso-scale dynamics of the molten alloy near the laser, known as the melt pool. For instance, the melt pool can directly contribute to the formation of undesirable porosity, residual stress, and surface roughness in the final part. Experimental in-situ monitoring of the three-dimensional melt pool physical fields is challenging, due to the short length and time scales involved in the process. Multi-physics simulation methods can describe the three-dimensional dynamics of the melt pool, but are computationally expensive at the mesh refinement required for accurate predictions of complex effects, such as the formation of keyhole porosity. Therefore, in this work, we develop a generative deep learning model based on the probabilistic diffusion framework to map low-fidelity, coarse-grained simulation information to the high-fidelity counterpart. By doing so, we bypass the computational expense of conducting multiple high-fidelity simulations for analysis by instead upscaling lightweight coarse mesh simulations. Specifically, we implement a 2-D diffusion model to spatially upscale cross-sections of the coarsely simulated melt pool to their high-fidelity equivalent. We demonstrate the preservation of key metrics of the melting process between the ground truth simulation data and the diffusion model output, such as the temperature field, the melt pool dimensions and the variability of the keyhole vapor cavity. Specifically, we predict the melt pool depth within 3 μm based on low-fidelity input data 4× coarser than the high-fidelity simulations, reducing analysis time by two orders of magnitude.

Cross-Modal Retrieval-enhanced code Summarization based on joint learning for retrieval and generation

Context:Code summarization refers to a task that automatically generates a natural language description of a code snippet to facilitate code comprehension. Existing methods have achieved satisfactory results by incorporating information retrieval into generative deep-learning models for reusing summaries of existing code. However, most of these existing methods employed non-learnable generic retrieval methods for content-based retrieval, resulting in a lack of diversity in the retrieved results during training, thereby making the model over-reliant on retrieved results and reducing the generative model’s ability to generalize to unknown samples. Objective:To address this issue, this paper introduces CMR-Sum: a novel Cross-Modal Retrieval-enhanced code Summarization framework based on joint learning for generation and retrieval tasks, where both two tasks are allowed to be optimized simultaneously. Method:Specifically, we use a cross-modal retrieval module to dynamically alter retrieval results during training, which enhances the diversity of the retrieved results and maintains a relative balance between the two tasks. Furthermore, in the summary generation phase, we employ a cross-attention mechanism to generate code summaries based on the alignment between retrieved and generated summaries. We conducted experiments on three real-world datasets, comparing the performance of our method with baseline models. Additionally, we performed extensive qualitative analysis. Result:Results from qualitative and quantitative experiments indicate that our approach effectively enhances the performance of code summarization. Our method outperforms both the generation-based and the retrieval-enhanced baselines. Further ablation experiments demonstrate the effectiveness of each component of our method. Results from sensitivity analysis experiments suggest that our approach achieves good performance without requiring extensive hyper-parameter search. Conclusion:The direction of utilizing retrieval-enhanced generation tasks shows great potential. It is essential to increase the diversity of retrieval results during the training process, which is crucial for improving the generality and the performance of the model.

Providing context: Extracting non-linear and dynamic temporal motifs from brain activity.

Approaches studying the dynamics of resting-state functional magnetic resonance imaging (rs-fMRI) activity often focus on time-resolved functional connectivity (tr-FC). While many approaches have been proposed, these typically focus on linear approaches like computing the linear correlation at a timestep or within a window. In this work, we propose to use a generative non-linear deep learning model, a disentangled variational autoencoder (DSVAE), that factorizes out window-specific (context) information from timestep-specific (local) information. This has the advantage of allowing our model to capture differences at multiple temporal scales. For the timestep-specific scale, which has higher temporal precision, we find significant differences between schizophrenia patients and control subjects in their temporal step distance through our model's latent space. We also find that window-specific embeddings, or as we refer to them, context embeddings, more accurately separate windows from schizophrenia patients and control subjects than the standard tr-FC approach. Moreover, we find that for individuals with schizophrenia, our model's context embedding space is significantly correlated with both age and symptom severity. Interestingly, patients appear to spend more time in three clusters, one closer to controls which shows increased visual-sensorimotor, cerebellar-subcortical, and reduced cerebellar-sensorimotor functional network connectivity (FNC), an intermediate station showing increased subcortical-sensorimotor FNC, and one that shows decreased visual-sensorimotor, decreased subcortical-sensorimotor, and increased visual-subcortical domains. We verify that our model captures features that are complementary to - but not the same as - standard tr-FC features. Our model can thus help broaden the neuroimaging toolset in analyzing fMRI dynamics and shows potential as an approach for finding psychiatric links that are more sensitive to individual and group characteristics.

Species-specific design of artificial promoters by transfer-learning based generative deep-learning model.

Native prokaryotic promoters share common sequence patterns,but are species dependent. For understudied species with limited data, it is challenging to predict the strength of existing promoters and generate novel promoters. Here, we developed PromoGen, a collection of nucleotide language models to generate species-specific functional promoters, across dozens of species in a data and parameter efficient way. Twenty-seven species-specific models in this collection were finetuned from the pretrained model which was trained on multi-species promoters. When systematically compared with native promoters, the Escherichia coli- and Bacillus subtilis-specific artificial PromoGen-generated promoters (PGPs) were demonstrated to hold all distribution patterns of native promoters. A regression model was developed to score generated either by PromoGen or by another competitive neural network, and the overall score of PGPs is higher. Encouraged by in silico analysis, we further experimentally characterized twenty-two B. subtilis PGPs, results showed that four of tested PGPs reached the strong promoter level while all were active. Furthermore, we developed a user-friendly website to generate species-specific promoters for 27 different species by PromoGen. This work presented an efficient deep-learning strategy for de novo species-specific promoter generation even with limited datasets, providing valuable promoter toolboxes especially for the metabolic engineering of understudied microorganisms.

Innovative methods for microplastic characterization and detection: Deep learning supported by photoacoustic imaging and automated pre-processing data

Plastic products' widespread applications and their non-biodegradable nature have resulted in the continuous accumulation of microplastic waste, emerging as a significant component of ecological environmental issues. In the field of microplastic detection, the intricate morphology poses challenges in achieving rapid visual characterization of microplastics. In this study, photoacoustic imaging technology is initially employed to capture high-resolution images of diverse microplastic samples. To address the limited dataset issue, an automated data processing pipeline is designed to obtain sample masks while effectively expanding the dataset size. Additionally, we propose Vqdp2, a generative deep learning model with multiple proxy tasks, for predicting six forms of microplastics data. By simultaneously constraining model parameters through two training modes, outstanding morphological category representations are achieved. The results demonstrate Vqdp2's excellent performance in classification accuracy and feature extraction by leveraging the advantages of multi-task training. This research is expected to be attractive for the detection classification and visual characterization of microplastics.

Exploiting a variational auto-encoder to represent the evolution of sudden stratospheric warmings

Sudden stratospheric warmings (SSWs) are the most dramatic events in the wintertime stratosphere. Such extreme events are characterized by substantial disruption to the stratospheric polar vortex, which can be categorized into displacement and splitting types depending on the morphology of the disrupted vortex. Moreover, SSWs are usually followed by anomalous tropospheric circulation regimes that are important for subseasonal-to-seasonal prediction. Thus, monitoring the genesis and evolution of SSWs is crucial and deserves further advancement. Despite several analysis methods that have been used to study the evolution of SSWs, the ability of deep learning methods has not yet been explored, mainly due to the relative scarcity of observed events. To overcome the limited observational sample size, we use data from historical simulations of the Whole Atmosphere Community Climate Model version 6 to identify thousands of simulated SSWs, and use their spatial patterns to train the deep learning model. We utilize a convolutional neural network combined with a variational auto-encoder (VAE)—a generative deep learning model—to construct a phase diagram that characterizes the SSW evolution. This approach not only allows us to create a latent space that encapsulates the essential features of the vortex structure during SSWs, but also offers new insights into its spatiotemporal evolution mapping onto the phase diagram. The constructed phase diagram depicts a continuous transition of the vortex pattern during SSWs. Notably, it provides a new perspective for discussing the evolutionary paths of SSWs: the VAE gives a better-reconstructed vortex morphology and more clearly organized vortex regimes for both displacement-type and split-type events than those obtained from principal component analysis. Our results provide an innovative phase diagram to portray the evolution of SSWs, in which particularly the splitting SSWs are better characterized. Our findings support the future use of deep learning techniques to study the underlying dynamics of extreme stratospheric vortex phenomena, and to establish a benchmark to evaluate model performance in simulating SSWs.

Accuracy of generative deep learning model for macular anatomy prediction from optical coherence tomography images in macular hole surgery

This study aims to propose a generative deep learning model (GDLM) based on a variational autoencoder that predicts macular optical coherence tomography (OCT) images following full-thickness macular hole (FTMH) surgery and evaluate its clinical accuracy. Preoperative and 6-month postoperative swept-source OCT data were collected from 150 patients with successfully closed FTMH using 6 × 6 mm2 macular volume scan datasets. Randomly selected and augmented 120,000 training and 5000 validation pairs of OCT images were used to train the GDLM. We assessed the accuracy and F1 score of concordance for neurosensory retinal areas, performed Bland–Altman analysis of foveolar height (FH) and mean foveal thickness (MFT), and predicted postoperative external limiting membrane (ELM) and ellipsoid zone (EZ) restoration accuracy between artificial intelligence (AI)-OCT and ground truth (GT)-OCT images. Accuracy and F1 scores were 94.7% and 0.891, respectively. Average FH (228.2 vs. 233.4 μm, P = 0.587) and MFT (271.4 vs. 273.3 μm, P = 0.819) were similar between AI- and GT-OCT images, within 30.0% differences of 95% limits of agreement. ELM and EZ recovery prediction accuracy was 88.0% and 92.0%, respectively. The proposed GDLM accurately predicted macular OCT images following FTMH surgery, aiding patient and surgeon understanding of postoperative macular features.

Development And Challenges of Generative Artificial Intelligence in Education and Art

Thanks to the rapid development of generative deep learning models, Artificial Intelligence Generated Content (AIGC) has attracted more and more research attention in recent years, which aims to learn models from massive data to generate relevant content based on input conditions. Different from traditional single-modal generation tasks that focus on content generation for a particular modality, such as image generation, text generation, or semantic generation, AIGC trains a single model that can simultaneously understand language, images, videos, audio, and more. AIGC marks the transition from traditional decision-based artificial intelligence to generative artificial intelligence, which has been widely applied in various fields. Focusing on the key technologies and representative applications of AIGC, this paper identifies several key technical challenges and controversies in the field. These include defects in cross-modal and multimodal generation, issues related to model stability and data consistency, privacy concerns, and questions about whether advanced generative models like ChatGPT can be considered general artificial intelligence (AGI). While this dissertation provides valuable insights into the revolution and challenge of generative AI in art and education, it acknowledges the sensitivity of generated content and the ethical dilemmas it may pose, and ownership rights for AI-generated works and the need for new intellectual property norms are subjects of ongoing discussion. To address the current technical bottlenecks in cross-modal and multimodal generation, future research aims to quantitatively analyze and compare existing models, proposing practical optimization strategies. With the rapid advancement of generative AI, we anticipate a transition from user-generated content (UGC) to artificial intelligence-generated content (AIGC) and, ultimately, a new era of human-computer co-creation with strong interactive potential in the near future.

An application of generative AI for knitted textile design in fashion

In recent years, artificial intelligence (AI) in the form of generative deep learning models have proliferated as a tool to facilitate or exhibit creativity across various design fields. When it comes to fashion design, existing applications of AI have more heavily addressed general fashion design elements, such as style, silhouette, colour, and pattern, and paid less attention to the underlying textile attributes. To address this gap, this study explores the effects of applying a generative deep learning model specifically towards the textile component of the fashion design process, by utilizing a Generative Adversarial Network (GAN) model to generate new images of knitted textile designs, which were then assessed based on their aesthetic quality in a qualitative survey with over 200 respondents. The results suggest that the generative deep learning (GAN) based method has the ability to produce new textile designs with creative qualities and practical utility that facilitate the fashion design process.

Data-Driven Based Characterization of Anisotropic Mechanical Properties for Cancellous Bone From Low-Resolution CT Images.

Exploring the anisotropic mechanical behavior of cancellous bone is crucial for in-vivo bone biomechanical analysis. However, it is challenging to characterize anisotropic mechanical behaviors under low-resolution (LR) clinical CT images due to a lack of microstructural information. The data-driven method proposed in this article accurately characterizes the anisotropic mechanical properties of cancellous bone from LR clinical CT images. The trabecular bone cubes of sheep are used to obtain a high-resolution (HR) micro-CT and an LR clinical CT image dataset. First, an auto-encoder model is trained using HR image data. Microstructural features are extracted by the encoder. A fast super-resolution (FSR) model is trained to map LR bone cubes to the features extracted from corresponding HR samples. The pretrained FSR model is used to convert LR clinical CT images to encoded microstructural features. The features are later used to predict target histomorphological parameters, anisotropic elastic tensors, and fabric tensors based on a fully connected neural network. The data-driven model accurately predicts the elastic tensor and fabric tensor of trabecular bones with LR CT images with 0.6 mm/pixel spatial resolution. It was verified that LR clinical CT images could generate microstructural information using a generative deep-learning model and an up-sampling operation. This study proves that clinical medical images of cancellous bone can be used for analysis of complex mechanical properties using a data-driven method, which is useful for real-time bone defect diagnosis and personalized bone prosthesis design in clinical application.

Dual contrastive learning based image-to-image translation of unstained skin tissue into virtually stained H&E images

Staining is a crucial step in histopathology that prepares tissue sections for microscopic examination. Hematoxylin and eosin (H&E) staining, also known as basic or routine staining, is used in 80% of histopathology slides worldwide. To enhance the histopathology workflow, recent research has focused on integrating generative artificial intelligence and deep learning models. These models have the potential to improve staining accuracy, reduce staining time, and minimize the use of hazardous chemicals, making histopathology a safer and more efficient field. In this study, we introduce a novel three-stage, dual contrastive learning-based, image-to-image generative (DCLGAN) model for virtually applying an "H&E stain" to unstained skin tissue images. The proposed model utilizes a unique learning setting comprising two pairs of generators and discriminators. By employing contrastive learning, our model maximizes the mutual information between traditional H&E-stained and virtually stained H&E patches. Our dataset consists of pairs of unstained and H&E-stained images, scanned with a brightfield microscope at 20 × magnification, providing a comprehensive set of training and testing images for evaluating the efficacy of our proposed model. Two metrics, Fréchet Inception Distance (FID) and Kernel Inception Distance (KID), were used to quantitatively evaluate virtual stained slides. Our analysis revealed that the average FID score between virtually stained and H&E-stained images (80.47) was considerably lower than that between unstained and virtually stained slides (342.01), and unstained and H&E stained (320.4) indicating a similarity virtual and H&E stains. Similarly, the mean KID score between H&E stained and virtually stained images (0.022) was significantly lower than the mean KID score between unstained and H&E stained (0.28) or unstained and virtually stained (0.31) images. In addition, a group of experienced dermatopathologists evaluated traditional and virtually stained images and demonstrated an average agreement of 78.8% and 90.2% for paired and single virtual stained image evaluations, respectively. Our study demonstrates that the proposed three-stage dual contrastive learning-based image-to-image generative model is effective in generating virtual stained images, as indicated by quantified parameters and grader evaluations. In addition, our findings suggest that GAN models have the potential to replace traditional H&E staining, which can reduce both time and environmental impact. This study highlights the promise of virtual staining as a viable alternative to traditional staining techniques in histopathology.

Development of a generative deep learning model to improve epiretinal membrane detection in fundus photography

BackgroundThe epiretinal membrane (ERM) is a common retinal disorder characterized by abnormal fibrocellular tissue at the vitreomacular interface. Most patients with ERM are asymptomatic at early stages. Therefore, screening for ERM will become increasingly important. Despite the high prevalence of ERM, few deep learning studies have investigated ERM detection in the color fundus photography (CFP) domain. In this study, we built a generative model to enhance ERM detection performance in the CFP.MethodsThis deep learning study retrospectively collected 302 ERM and 1,250 healthy CFP data points from a healthcare center. The generative model using StyleGAN2 was trained using single-center data. EfficientNetB0 with StyleGAN2-based augmentation was validated using independent internal single-center data and external datasets. We randomly assigned healthcare center data to the development (80%) and internal validation (20%) datasets. Data from two publicly accessible sources were used as external validation datasets.ResultsStyleGAN2 facilitated realistic CFP synthesis with the characteristic cellophane reflex features of the ERM. The proposed method with StyleGAN2-based augmentation outperformed the typical transfer learning without a generative adversarial network. The proposed model achieved an area under the receiver operating characteristic (AUC) curve of 0.926 for internal validation. AUCs of 0.951 and 0.914 were obtained for the two external validation datasets. Compared with the deep learning model without augmentation, StyleGAN2-based augmentation improved the detection performance and contributed to the focus on the location of the ERM.ConclusionsWe proposed an ERM detection model by synthesizing realistic CFP images with the pathological features of ERM through generative deep learning. We believe that our deep learning framework will help achieve a more accurate detection of ERM in a limited data setting.

P505 Accelerating Crohn's Disease clinical trials using generative artificial intelligence

Abstract Background At Unlearn.AI, we expedite decision making and address challenges of lengthy and costly Crohn’s Disease (CD) trial enrollment by introducing deep learning models that create digital twins of trial participants, which are comprehensive forecasts of health outcomes. These generative predictions are incorporated into standard linear and logistic regression analyses, in addition to more advanced methods, in clinical trials. Our TwinRCT technology, which aligns with regulatory guidance, augments standard RCTs for easy integration into clinical programs. We demonstrate the expected benefit of these solutions in a subset of IBD Plexus data that were not previously involved in the training of the generative model, showcasing its potential to transform future CD trials. Methods We developed a generative deep learning model using data from approximately 4,500 Inflammatory Bowel Disorder (IBD) participants in observational studies and randomized controlled trial (RCT) control arms from the IBD Plexus program of the Crohn’s & Colitis Foundation. The trained model was used to forecast, entirely from each participant’s baseline data, the control outcomes of participants in a held out set of 300 participants. The resultant forecasts, or digital twins, represent comprehensive longitudinal trajectories across a range of outcomes, laboratory measures, and vital signs. We evaluated the expected benefits of incorporating the digital twins in linear and logistic regression models, analyzing the change from baseline to follow-up (Weeks 12 through 52) on sCDAI total score, change in daily bowel movement count, change in abdominal pain score and proportion of patients achieving clinical remission (sCDAI < 150; sCDAI is very closely related to CDAI). Results Digital twins provided substantial attainable benefit across outcomes and time points, enabling up to 21% overall sample size reduction or 34% reduction in the control arm only. Conclusion TwinRCTs offer a strategic enhancement to traditional RCTs by minimizing the required sample size needed to identify significant treatment effects. Aligning with existing FDA and EMA guidelines, TwinRCTs can be effectively utilized in Phase 2 and Phase 3 CD studies for faster decision making, as well as reduced timelines and costs, without increasing regulatory risk over standard approaches.

A feasibility study of applying generative deep learning models for map labeling

ABSTRACT The automation of map labeling is an ongoing research challenge. Currently, the map labeling algorithms are based on rules defined by experts for optimizing the placement of the text labels on maps. In this paper, we investigate the feasibility of using well-labeled map samples as a source of knowledge for automating the labeling process. The basic idea is to train deep learning models, specifically the generative models CycleGAN and Pix2Pix, on a large number of map examples. Then, the trained models are used to predict good locations of the labels given unlabeled raster maps. We compare the results obtained by the deep learning models to manual map labeling and a state-of-the-art optimization-based labeling method. A quantitative evaluation is performed in terms of legibility, association and map readability as well as a visual evaluation performed by three professional cartographers. The evaluation indicates that the deep learning models are capable of finding appropriate positions for the labels, but that they, in this implementation, are not well suited for selecting the labels to show and to determine the size of the labels. The result provides valuable insights into the current capabilities of generative models for such task, while also identifying the key challenges that will shape future research directions.

Super-Resolution of Radargrams With a Generative Deep Learning Model

Super-Resolution of Radargrams With a Generative Deep Learning Model

Identifying compound weather drivers of forest biomass loss with generative deep learning

Abstract Globally, forests are net carbon sinks that partly mitigates anthropogenic climate change. However, there is evidence of increasing weather-induced tree mortality, which needs to be better understood to improve forest management under future climate conditions. Disentangling drivers of tree mortality is challenging because of their interacting behavior over multiple temporal scales. In this study, we take a data-driven approach to the problem. We generate hourly temperate weather data using a stochastic weather generator to simulate 160,000 years of beech, pine, and spruce forest dynamics with a forest gap model. These data are used to train a generative deep learning model (a modified variational autoencoder) to learn representations of three-year-long monthly weather conditions (precipitation, temperature, and solar radiation) in an unsupervised way. We then associate these weather representations with years of high biomass loss in the forests and derive weather prototypes associated with such years. The identified prototype weather conditions are associated with 5–22% higher median biomass loss compared to the median of all samples, depending on the forest type and the prototype. When prototype weather conditions co-occur, these numbers increase to 10–25%. Our research illustrates how generative deep learning can discover compounding weather patterns associated with extreme impacts.

Estimating the environmental impact of Generative-AI services using an LCA-based methodology

As digital services are increasingly being deployed and used in a variety of domains, the environmental impact of Information and Communication Technologies (ICTs) is a matter of concern. Artificial intelligence is driving some of this growth but its environmental cost remains scarcely studied. A recent trend in large-scale generative models such as ChatGPT has especially drawn attention since their training requires intensive use of a massive number of specialized computing resources. The inference of those models is made accessible on the web as services, and using them additionally mobilizes end-user terminals, networks, and data centers. Therefore, those services contribute to global warming, worsen metal scarcity, and increase energy consumption. This work proposes an LCA-based methodology for a multi-criteria evaluation of the environmental impact of generative AI services, considering embodied and usage costs of all the resources required for training models, inferring from them, and hosting them online. We illustrate our methodology with Stable Diffusion as a service, an open-source text-to-image generative deep-learning model accessible online. This use case is based on an experimental observation of Stable Diffusion training and inference energy consumption. Through a sensitivity analysis, various scenarios estimating the influence of usage intensity on the impact sources are explored.

Research on Pricing Methods of Convertible Bonds Based on Deep Learning GAN Models

This paper proposes two data-driven models (including LSTM pricing model, WGAN pricing model) and an improved model of LSM based on GAN to analyze the pricing of convertible bonds. In addition, the LSM model with higher precision in traditional pricing model is selected for comparative study with other pricing models. It is found that the traditional LSM pricing model has a large error in the first-day pricing, and the pricing function needs to be further improved. Among the four pricing models, LSTM pricing model and WGAN pricing model have the best pricing effect. The WGAN pricing model is better than the LSTM pricing model (0.21%), and the LSM improved model (1.17%) is better than the traditional LSM model (2.26%). Applying the generative deep learning model GAN to the pricing of convertible bonds can circumvent the harsh preconditions of assumptions, and significantly improve the pricing effect of the traditional model. The scope of application of each model is different. Therefore, this paper proves the feasibility of the GAN model applied to the pricing of convertible bonds, and enriches the pricing function of derivatives in the financial field.

Generating bulk RNA-Seq gene expression data based on generative deep learning models and utilizing it for data augmentation

Large-scale high-throughput transcriptome sequencing data holds significant value in biomedical research. However, practical challenges such as difficulty in sample acquisition often limit the availability of large sample sizes, leading to decreased reliability of the analysis results. In practice, generative deep learning models, such as Generative Adversarial Networks (GANs) and Diffusion Models (DMs), have been proven to generate realistic data and may be used to solve this promblem. In this study, we utilized bulk RNA-Seq gene expression data to construct different generative models with two data preprocessing methods: Min-Max-GAN, Z-Score-GAN, Min-Max-DM, and Z-Score-DM. We demonstrated that the generated data from the Min-Max-GAN model exhibited high similarity to real data, surpassing the performance of the other models significantly. Furthermore, we trained the models on the largest dataset available to date, achieving MMD (Maximum Mean Discrepancy) of 0.030 and 0.033 on the training and independent datasets, respectively. Through SHAP (SHapley Additive exPlanations) explanations of our generative model, we also enhanced our model's credibility. Finally, we applied the generated data to data augmentation and observed a significant improvement in the performance of classification models. In summary, this study establishes a GAN-based approach for generating bulk RNA-Seq gene expression data, which contributes to enhancing the performance and reliability of downstream tasks in high-throughput transcriptome analysis.

Multi-objective generative design of three-dimensional material structures

Generative design for materials has recently gained significant attention due to the rapid evolution of generative deep learning models. There have been a few successful generative design demonstrations of molecular-level structures with the help of graph neural networks. However, in the realm of macroscale material structures, most of the works are targeting two-dimensional, ungoverned structure generations. Hindered by the complexity of 3D structures, it is hard to extract customized structures with multiple desired properties from a large, unexplored design space. Here we report a novel framework, a multi-objective driven Wasserstein generative adversarial network (WGAN), to implement inverse designs of 3D structures according to given geometrical, structural, and mechanical requirements. Our framework consists of a WGAN-based network that generates 3D structures possessing geometrical and structural features learned from the target dataset. Besides, multiple objectives are introduced to our framework for the control of mechanical property and isotropy of the structures. An accurate surrogate model is incorporated into the framework to perform efficient prediction on the properties of generated structures in training iterations. With multiple objectives combined by their weight and the 3D WGAN acting as a soft constraint to regulate features that are hard to define by the traditional method, our framework has proven to be capable of tuning the properties of the generated structures in multiple aspects while keeping the selected structural features. The feasibility of a small dataset and the scalability of the objectives of other properties make our work an effective approach to provide fast and automated structure designs for various functional materials.