Large-scale scientific simulations require significant resources to generate high-resolution time-varying data (TVD). While super-resolution is an efficient post-processing strategy to reduce costs, existing methods rely on a large amount of HR training data, limiting their applicability to diverse simulation scenarios. To address this constraint, we proposed CD-TVD, a novel framework that combines contrastive learning and an improved diffusion-based super-resolution model to achieve accurate 3D super-resolution from limited time-step high-resolution data. During pre-training on historical simulation data, the contrastive encoder and diffusion super-resolution modules learn degradation patterns and detailed features of high-resolution and low-resolution samples. In the training phase, the improved diffusion model with a local attention mechanism is fine-tuned using only one newly generated high-resolution timestep, leveraging the degradation knowledge learned by the encoder. This design minimizes the reliance on large-scale high-resolution datasets while maintaining the capability to recover fine-grained details. Experimental results on fluid and atmospheric simulation datasets confirm that CD-TVD delivers accurate and resource-efficient 3D super-resolution, marking a significant advancement in data augmentation for large-scale scientific simulations. The code is available at https://github.com/Xin-Gao-private/CD-TVD.

To study complex real-world phenomena using computer simulations, scientists often rely on ensemble datasets generated from multiple simulation runs with varying parameter configurations. This process can produce ensemble datasets with many members, making traditional data analysis pipelines impractical due to limited I/O bandwidth and disk capacity. Distribution-based data representations have been proposed as a promising solution. Processing data in situ to generate compact distribution-based representations not only alleviates the challenges of limited I/O bandwidth and disk capacity but also enables uncertainty quantification, thus mitigating the risk of misinterpretation. Nevertheless, distribution-based methods inherently sacrifice spatial information of data samples within the distribution, potentially reducing precision in the data analysis pipeline. To address this issue, we introduce a deep learning model to reconstruct data volume from the distribution representation. Instead of using a model that predicts a data block directly from its distribution representation, we propose a deep learning model based on the Sinkhorn operator and Gumbel trick that learns to map samples drawn from a distribution to spatial locations within the block. The deep learning model can support high-quality downstream data analysis and visualization, provide point-wise uncertainty quantification, and guarantee the distribution of the reconstructed data block follows the block's distribution representation.

Driven by advances in supercomputing, the scale of scientific simulation data has grown dramatically. In fields such as cosmology, particle data have become a common representation, with state-of-the-art simulations now exceeding the trillion-particle mark. Consequently, the challenge of visually analyzing such massive datasets has become increasingly urgent. The traditional visual analysis workflow typically follows a "compression $\rightarrow$→ storage $\rightarrow$→ reconstruction $\rightarrow$→ visualization" pipeline. However, this process is hampered by an extremely time-consuming reconstruction stage, which severely impedes real-time interactive visualization. Moreover, in multi-time-step analyses, the enormous volume of reconstructed data creates significant I/O bottlenecks. In this work, we draw inspiration from 3D Gaussian splatting and compress the simulation data using Gaussian Mixture Models (GMMs), treating the resulting Gaussian kernels as fundamental rendering primitives. Our method renders billion-scale particles for each timestep in approximately 32 ms, requiring only 645 MB of GPU memory per timestep - nearly 20× smaller than the original 12 GB raw data. This eliminates costly reconstruction, accelerates the visual analysis pipeline, and overcomes I/O bottlenecks in multi-time-step analysis. Extensive experiments and comparisons across multiple datasets validate the effectiveness of our method.

Meteorological experts achieved promising results in improving the effectiveness of neural weather networks in precipitation forecasting by incorporating precipitation-optimized objectives into the loss function. However, since neural networks are not based on explicit physical processes, meteorological experts may lack understanding and trust in the model’s predictions, and they cannot perform bias correction by adjusting physical parameters. Additionally, precipitation events are highly imbalanced, with heavy rainfall events being relatively rare but of greater importance. Therefore, traditional metrics for evaluating deep models are insufficient to fully assess precipitation forecasting performance. In this paper, we present a neural weather network visual analysis system designed to help domain experts understand and comprehensively compare the impact of different loss functions on neural networks. We customize the model evaluation process based on the characteristics of precipitation forecasting tasks and provide a reference for bias correction using historically similar data. To validate our approach, we perform two case studies using real-world reanalysis datasets, with feedback from domain experts further confirming its effectiveness.

Abstract Particle advection is a foundational algorithm for analyzing a flow field. The commonly used Parallelization‐Over‐Data (POD) strategy for particle advection can become slow and inefficient when there are unbalanced workloads, which are particularly prevalent in in situ workflows. In this work, we present an in situ workflow containing workload estimation for block assignment and duplication in a parallelization‐over‐data algorithm. With tightly coupled workload estimation and load‐balanced block assignment strategy, our workflow offers a considerable improvement over the traditional round‐robin block assignment strategy. Our experiments demonstrate that particle advection is up to 3X faster and associated workflow saves approximately 30% of execution time after adopting strategies presented in this work.

Numerical simulation serves as a cornerstone in scientific modeling, yet the process of fine-tuning simulation parameters poses significant challenges. Conventionally, parameter adjustment relies on extensive numerical simulations, data analysis, and expert insights, resulting in substantial computational costs and low efficiency. The emergence of deep learning in recent years has provided promising avenues for more efficient exploration of parameter spaces. However, existing approaches often lack intuitive methods for precise parameter adjustment and optimization. To tackle these challenges, we introduce ParamsDrag, a model that facilitates parameter space exploration through direct interaction with visualizations. Inspired by DragGAN, our ParamsDrag model operates in three steps. First, the generative component of ParamsDrag generates visualizations based on the input simulation parameters. Second, by directly dragging structure-related features in the visualizations, users can intuitively understand the controlling effect of different parameters. Third, with the understanding from the earlier step, users can steer ParamsDrag to produce dynamic visual outcomes. Through experiments conducted on real-world simulations and comparisons with state-of-the-art deep learning-based approaches, we demonstrate the efficacy of our solution.

Deep reinforcement learning (DRL) aims to train software agents that can understand environments and learn effective strategies, and has achieved significant breakthroughs in performance and capabilities, particularly in areas such as Go, Atari games, and autonomous vehicles. Unlike traditional deep learning, the goals of reinforcement learning can be more abstract and require careful modification of reward functions. The training process involves unstructured sequential data, which can be difficult for human experts to analyze and gain insights from. To address this challenge, we propose SampleViz, a visual analytics system that enables flexible interaction between human experts and DRL sequence data, allowing for the extraction of crucial concepts from massive amounts of data and their provision to the agent. SampleViz transforms the tedious task of modifying reward functions and policy debugging into an engaging concept exploration process, allowing for the efficient integration of human expertise with automatic sampling algorithms for effective model improvement. Through case studies and expert feedback, we demonstrate that SampleViz can effectively assist experts in concept extraction and model improvement, and enables the incorporation of interpretability and human-in-the-loop concepts into DRL policy settings.

The past decade has witnessed the superior power of deep neural networks (DNNs) in applications across various domains. However, training a high-quality DNN remains a non-trivial task due to its massive number of parameters. Visualization has shown great potential in addressing this situation, as evidenced by numerous recent visualization works that aid in DNN training and interpretation. These works commonly employ a strategy of logging training-related data and conducting post-hoc analysis. Based on the results of offline analysis, the model can be further trained or fine-tuned. This strategy, however, does not cope with the increasing complexity of DNNs, because (1) the time-series data collected over the training are usually too large to be stored entirely; (2) the huge I/O overhead significantly impacts the training efficiency; (3) post-hoc analysis does not allow rapid human-interventions (e.g., stop training with improper hyper-parameter settings to save computational resources). To address these challenges, we propose an in-situ visualization and analysis framework for the training of DNNs. Specifically, we employ feature extraction algorithms to reduce the size of training-related data in-situ and use the reduced data for real-time visual analytics. The states of model training are disclosed to model designers in real-time, enabling human interventions on demand to steer the training. Through concrete case studies, we demonstrate how our in-situ framework helps deep learning experts optimize DNNs and improve their analysis efficiency.

N-body numerical simulation is an important tool in astronomy. Scientists used this method to simulate the formation of structure of the universe, which is key to understanding how the universe formed. As research on this subject further develops, astronomers require a more precise method that enables expansion of the simulation and an increase in the number of simulation particles. However, retaining all temporal information is infeasible due to a lack of computer storage. In the circumstances, astronomers reserve temporal data at intervals, merging rough and baffling animations of universal evolution. In this study, we propose a deep-learning-assisted interpolation application to analyze the structure formation of the universe. First, we evaluate the feasibility of applying interpolation to generate an animation of the universal evolution through an experiment. Then, we demonstrate the superiority of deep convolutional neural network (DCNN) method by comparing its quality and performance with the actual results together with the results generated by other popular interpolation algorithms. In addition, we present PRSVis, an interactive visual analytics system that supports global volume rendering, local area magnification, and temporal animation generation. PRSVis allows users to visualize a global volume rendering, interactively select one cubic region from the rendering and intelligently produce a time-series animation of the high-resolution region using the deep-learning-assisted method. In summary, we propose an interactive visual system, integrated with the DCNN interpolation method that is validated through experiments, to help scientists easily understand the evolution of the particle region structure.

Convolutional neural networks (CNNs) have demonstrated extraordinarily good performance in many computer vision tasks. The increasing size of CNN models, however, prevents them from being widely deployed to devices with limited computational resources, e.g., mobile/embedded devices. The emerging topic of model pruning strives to address this problem by removing less important neurons and fine-tuning the pruned networks to minimize the accuracy loss. Nevertheless, existing automated pruning solutions often rely on a numerical threshold of the pruning criteria, lacking the flexibility to optimally balance the trade-off between efficiency and accuracy. Moreover, the complicated interplay between the stages of neuron pruning and model fine-tuning makes this process opaque, and therefore becomes difficult to optimize. In this paper, we address these challenges through a visual analytics approach, named CNNPruner. It considers the importance of convolutional filters through both instability and sensitivity, and allows users to interactively create pruning plans according to a desired goal on model size or accuracy. Also, CNNPruner integrates state-of-the-art filter visualization techniques to help users understand the roles that different filters played and refine their pruning plans. Through comprehensive case studies on CNNs with real-world sizes, we validate the effectiveness of CNNPruner.

Cosmological N-body simulation is an important tool for scientists to study the evolution of the universe. With the increase of computing power, billions of particles of high space-time fidelity can be simulated by supercomputers. However, limited computer storage can only hold a small subset of the simulation output for analysis, which makes the understanding of the underlying cosmological phenomena difficult. To alleviate the problem, we design an in-situ data reduction method for large-scale unstructured particle data. During the data generation phase, we use a combined k-dimensional partitioning and Gaussian mixture model approach to reduce the data by utilizing probability distributions. We offer a model evaluation criterion to examine the quality of the probabilistic distribution models, which allows us to identify and improve low-quality models. After the in-situ processing, the particle data size is greatly reduced, which satisfies the requirements from the domain experts. By comparing the astronomical attributes and visualizations of the reconstructed data with the raw data, we demonstrate the effectiveness of our in-situ particle data reduction technique.