利用深度學習輔助大規模宇宙模擬的視覺化參數空間分析
No Thumbnail Available
Date
2021
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
宇宙學家經常建立數學模擬模型來研究觀察到的宇宙。隨著計算能力的大幅提升,模擬可以進行更多的細節生成,並減少觀察到的宇宙與模擬結果之間的差異。然而,高保真的模擬運算耗時且給分析帶來不便,特別是當涉及大量的輸入參數組合時。因此,選擇一個能夠滿足分析任務需要的輸入參數組合就成為分析過程的重要部分之一。在這項工作中,我們提出了一個交互式視覺系統。它可以幫助用戶有效地理解大規模宇宙學數據的參數空間,並進一步發現有價值的模擬輸入參數組合。我們的系統利用基於 GAN 的代理模型來重建模擬輸出,而無需運行每個實例都要花費大量時間的原始昂貴模擬模型。我們還提取了基於深度神經網絡的代理模型學習到的信息,以促進參數空間的探索。例如,隱藏層的輸出用於估計輸入參數配置之間的模擬輸出相似性。模擬參數的敏感性是使用反向傳播從代理模型中估計出來的。我們通過多個案例研究證明了我們系統的有效性,包括發現有價值的模擬輸入參數配置和子區域分析。
Cosmologists often built a mathematics simulation model to study the observed universe. With the momentous improvements in computing power, the simulation can conduct more details and alleviate the discrepancy between the observed universe and simulation result. However, the high-fidelity simulation is time-consuming and brings inconvenience to the analysis. Especially the simulation often involves a large number of simulation input parameter combinations. Therefore, selecting a combination of input parameters that can meet the needs of the analysis task has become an important part of the analysis process. In this work, we propose an interactive visual system, which helps users efficiently understand the parameter space of large-scale cosmology data and further discover the valuable simulation input parameter combinations. Our system utilizes the GAN-based surrogate models to reconstruct the simulation outputs without running the original expensive simulation that costs lots of time per instance. We also extract information learned by the deep neural network-based surrogate models to facilitate the parameter space exploration. For example, the outputs of the hidden layer are used to estimate the simulation output similarity among input parameter configurations. The sensitivities of the simulation parameters are estimated from the surrogate models using back-propagation. We demonstrate the effectiveness of our system with multiple case studies, including the discovery of valuable simulation input parameter configuration and sub-region analysis.
Cosmologists often built a mathematics simulation model to study the observed universe. With the momentous improvements in computing power, the simulation can conduct more details and alleviate the discrepancy between the observed universe and simulation result. However, the high-fidelity simulation is time-consuming and brings inconvenience to the analysis. Especially the simulation often involves a large number of simulation input parameter combinations. Therefore, selecting a combination of input parameters that can meet the needs of the analysis task has become an important part of the analysis process. In this work, we propose an interactive visual system, which helps users efficiently understand the parameter space of large-scale cosmology data and further discover the valuable simulation input parameter combinations. Our system utilizes the GAN-based surrogate models to reconstruct the simulation outputs without running the original expensive simulation that costs lots of time per instance. We also extract information learned by the deep neural network-based surrogate models to facilitate the parameter space exploration. For example, the outputs of the hidden layer are used to estimate the simulation output similarity among input parameter configurations. The sensitivities of the simulation parameters are estimated from the surrogate models using back-propagation. We demonstrate the effectiveness of our system with multiple case studies, including the discovery of valuable simulation input parameter configuration and sub-region analysis.
Description
Keywords
三維資料重建, 參數空間探索, 深度學習, 交互式視覺化系統, 代理模型, 宇宙學, 3D-reconstruction, Cosmology, deep learning, Surrogate modeling, Visual analysis, Parameter analysis