Research

Trustworthy Deep Learning for SciVis

Deep learning has been widely used to help scientific simulation and data processing. For example, surrogate models can replace expensive simulation runs by predicting the simulation outputs directly. Super-resolution methods can recover high-resolution data from low-resolution simulation data proxies, which reduces the cost of running simulations or directly stores high-resolution datasets. However, deep learning models work as black boxes. If we do not know when and where a model may fail, it can produce incorrect results without any warning, and these results may mislead the subsequent analysis. In ViDA Lab, we develop deep learning methods that not only generate and reconstruct scientific data with high quality but also estimate how confident the model is about its outputs. The estimated confidence can then be used in uncertainty visualization and data analysis. This allows scientists to know which parts of the results are reliable and to make decisions with more confidence when deep learning is used to assist their analysis.

Publicaionts:

• Han Huang, Zheng-Han Huang, Nathania Josephine, Guan Li, and Ko-Chih Wang. Trustworthy Deep Learning-Assisted Visualization and Analysis for Distribution-Based Ensemble Scientific Data Summarization. IEEE Transactions on Visualization and Computer Graphics (TVCG), 2026.

Distribution-based Analysis of Large-scale Scientific Data

Because of the advances in the computational power of modern supercomputers, scientists can simulate physical phenomena with higher spatial and temporal resolution and more ensemble members to support higher-quality scientific analysis. However, the large size of the output data causes long I/O time and storage problems. Naively downsampling the data before storage not only defeats the purpose of running high-resolution simulations but also produces artifacts that could mislead scientific data analysis and visualization. Distribution-based approaches have been recognized as a promising data representation because they can reduce the data size while quantifying the information loss, which supports uncertainty quantification and avoids misleading data analysis. In ViDA Lab, we focus on developing advanced distribution-based data representations that support high-quality visualization for high-resolution, multivariate, and ensemble simulation data.

Publicaionts:

• Zhi-Rong Lin, An-Lung Hsiao, Guan Li and Ko-Chih Wang. Distribution-based Time-varying Ensemble Scientific Data Reduction for Uncertainty Visualization and Analysis. Journal of Information Science and Engineering 40.2 (2024): 397-419.

• Yang, Hao-Yi, Zhi-Rong Lin and Ko-Chih Wang, Efficient and Portable Distribution Modeling for Large-Scale Scientific Data Processing with Data-Parallel Primitives, Algorithms 14.10 (2021): 285.

• Ko-Chih Wang, Tzu-Hsuan Wei, Shareef Naeem, and Han-Wei Shen, Ray-based Exploration of Large Time-varying Volume Data Using Per-ray Proxy Distributions, IEEE Transactions on Visualization and Computer Graphics, 2019. 

• Ko-Chih Wang, Kewei Lu, Tzu-Hsuan Wei, Naeem Shareef and Han-Wei Shen, Statistical visualization and analysis of large data using a value-based spatial distribution. IEEE PacificVis 2017: 161-170.

Visual Analytics for Explainable AI (XAI)

Since the resurgence of neural networks in the early 2010s, deep learning has become the most popular approach in many applications and domains, such as image recognition, natural language processing, and scientific research. Although deep learning has shown superior ability over many traditional approaches in many problems, a deep learning model works like a black box. Unlike traditional algorithms whose rules are explicitly designed by humans, it is extremely hard to understand how a deep learning model makes an inference and to determine whether the model is reliable. Therefore, it is hard for humans to predict when and where the model could fail, and this lack of trust becomes a barrier to applying deep learning models to critical applications, such as autonomous driving and weather forecasting. XAI has become a popular research direction that focuses on explicitly explaining what information the model learns, which information the model relies on to make a decision, and when and where the model is more likely to fail. In ViDA Lab, we focus on developing visual analytics techniques that assist machine learning experts and scientists in opening the black box of their models. Our techniques help them interpret, diagnose, and debug their models, which allows them to improve the models and determine whether the models are reliable enough for critical applications.

Publicaionts:

• Pei-Chen Chang, Zheng-Han Huang, Wan-Ling Tseng, Yi-Chi Wang, Hsin-Chien Liang, Cheng-Ta Chen, and Ko-Chih Wang. A Visual Analytics Approach to Exploring Regional Physical Processes Reflected in Generative Climate Downscaling Models. Information Visualization, 2026.

• Jian-Lin Wu, Pei-Chen Chang, Chao Wang, and Ko-Chih Wang. ATICVis: A Visual Analytics System for Asymmetric Transformer Models Interpretation and Comparison. Applied Sciences 13, no. 3: 1595 (2023). 

• Syu-Ting Deng, Cheng-Jun Tsai, Pei-Chen Chang and Ko-Chih Wang. Visual Interpretation for Contextualized Word Representation. Visualization Meets AI Workshop 2023.

• Pei-Shan Lo, Jian-Lin Wu, Syu-Ting Deng and Ko-Chih Wang, CNERVis: a visual diagnosis tool for Chinese named entity recognition. Journal of Visualization (2021): 1-17.  

Parameter Space Exploration of Ensemble Simulations

A computer simulation often takes a set of parameters as input to generate the corresponding data for scientific studies. Running a simulation with many different parameter settings produces an ensemble of simulation outputs. The relationship between the input parameters and the simulation output often has significant scientific meaning, such as how many degrees of ocean warming lead to which global climate patterns, or which value of the Hubble constant can reproduce the evolution of the universe we observe. Therefore, one major task in the simulation data analysis pipeline is to find the parameter settings that produce the outputs scientists are interested in. However, this task is often time-consuming because scientists need to guess a parameter setting based on their experience, wait for the simulation to finish, analyze the output, and repeat this procedure until they find the desired result. In ViDA Lab, we are interested in developing tools and new algorithms to support efficient parameter space exploration, which allow users to intuitively and quickly converge to the simulation parameters that generate the data they are interested in.

Publicaionts:

• Guan Li, Yang Liu, Guihua Shan, Shiyu Cheng, Weiqun Cao, Junpeng Wang, and Ko-Chih Wang. ParamsDrag: Interactive Parameter Space Exploration via Image-Space Dragging. IEEE Transactions on Visualization and Computer Graphics (TVCG), 2025.

• Cheng Sun and Ko-Chih Wang, DLA-VPS: Deep Learning Assisted Visual Parameter Space Analysis of Cosmological Simulations. IEEE Computer Graphics and Applications (2022).

• Wenbin He, Junpeng Wang, Hanqi Guo, Ko-Chih Wang, Han-Wei Shen, Mukund Raj, Youssef S. G. Nashed, and Tom Peterka, InSituNet: Deep Image Synthesis for Parameter Space Exploration of Ensemble Simulations, IEEE Transactions on Visualization and Computer Graphics, 2019. (IEEE VIS 2019  (SciVis) Best Paper Award)

• Subhashis Hazarika, Haoyu Li, Ko-Chih Wang, Han-Wei Shen, and Ching-Shan Chou, NNVA: Neural Network Assisted Visual Analysis of Yeast Cell Polarization Simulation, IEEE Transactions on Visualization and Computer Graphics, 2019. (IEEE VIS 2019  (VAST) Best Paper Honorable Mention) 

• Ko-Chih Wang, Jiayi Xu, Jonathan Woodring and Han-Wei Shen, Statistical Super Resolution for Data Analysis and Visualization of Large Scale Cosmological Simulations. IEEE PacificVis 2019. 

LLM x Visualization

With the fast development of large language models (LLMs) and LLM agents, they have been used to assist many tasks in our daily life and work. However, general-purpose LLMs cannot always be directly used in specific scenarios. For example, in education, directly applying an LLM may give students too much or irrelevant guidance and lower their learning efficiency. In addition, when LLM agents are used in coding and scientific data analysis, they reduce users’ effort but also reduce users’ understanding of the analysis pipeline, which makes it harder for users to find mistakes or dig deeper into the data. In ViDA Lab, we design frameworks that constrain how LLMs work in a target scenario, where the goal is not only to complete the task correctly but also to give users proper feedback during the process, so that users can easily get the information they need about what the LLM has done, rather than reading through every detail or only checking the final result. 

Publicaionts:

• Yu-Ling Wang, Nathania Josephine, and Ko-Chih Wang. NarratorVis: Automated Context-Aware Visual Data Story Generation Using Rule-Based Approach and Large Language Model. Information Visualization, 2026.

• Nathania Josephine, Zheng-Han  Huang, Shang-Chun Chang, and Ko-Chih Wang. Automated Narrative Generation from Tabular Data using a Rule-Based and LLM Fusion Approach. The 10th International Conference on Science and Technology (ICST) 2024. (Best Paper Award) 

Deep Learning for Weather and Climate Applications

Computer simulations play a very important role in weather forecasting and climate studies. They are used in many steps of the analysis pipeline, such as global numerical weather prediction, regional downscaling to produce high-resolution local information, and long-term climate projections under different future scenarios. However, these simulations often require long computation time and consume a large amount of energy. The long computation time limits how many simulations scientists can run, for example, the number of ensemble members, which is important for estimating forecast uncertainty. In addition, the energy consumption is also an important sustainability issue. The weather and climate domain has  accumulated a large amount of data, such as observations, reanalysis datasets, and outputs of previous simulations, which makes this domain a natural fit for deep learning because deep learning requires a large amount of training data. In ViDA Lab, we closely collaborate with atmospheric scientists and agencies to develop deep learning surrogate models that support fast and energy-efficient prediction as an alternative to classic simulations, and apply these models to real-world problems.
Publicaionts:

• Yi-Chi Wang, Chia-Hao Chiang, Chiung-Jui Su, Ko-Chih Wang, Wan-Ling Tseng, and Hsin-Chien Liang. Improving Daily Rainfall Downscaling over Taiwan's Complex Terrain with Surface-wind-informed Deep Learning Model. International Journal of Climatology, 2026.