This conference paper was submitted for presentation at the NAFEMS World Congress 2025, held in Salzburg, Austria from May 19–22, 2025.

Abstract

The term 'œDigital Twin' may correspond to many different definitions, from the scientific modeling of a physical phenomenon to an enhanced virtual supervision system of a complex industrial setup. We consider in this presentation the specific flavor where we want to gather data from an actual system, usually through sensors, feed a numerical model with the real time data, visualize the result, make predictions with derived quantities, and eventually influence the actual system with a feedback loop. However, this process is not new. It'™s very similar to the supervision process of industrial installations. The main difference is the recent technological breakthrough to manage very large data as input, throughput and output of the models, which leverages novel artificial intelligence based techniques. Processing large data also means using High Performance Computing capabilities to tackle data that a classic workstation will not be able to open and process. Visualization is a powerful tool for Digital Twins. For example, while the Digital Twin is running, being able to extract meaningful information from sensor data and to display them is a natural way to monitor the actual system. But scientific visualization could also be successfully used while building the Digital Twin and monitoring the correctness of the building process. When data is massive, specialized software with client/server architecture should be used for data analysis on high performance computers. These scalability challenges with massive data have been solved for many decades now and by combining these tools in a Digital Twin, the users can interact with large data living in a computer cluster while keeping a responsive user experience. Under the hood, the users would interact remotely with the data and trigger new analysis which will be computed on the cluster where the massive data is gathered. Only the result of the analysis and visualization will be streamed to the users, like images or videos. There is a large ecosystem around these scientific visualization tools, including bridges to web front-end, Virtual Reality devices, Python AI libraries, or advanced rendering back-end. Some real use cases are climate predictions made by DKRZ, where meaningful features are extracted from Petabytes of data, or Weather prediction scenarios from The Cyprus Institute. Another example from academic studies is the CALM-AA project where aero-acoustics measurements are visualized through Virtual Reality headsets with live interaction with the data. We also mention the VESTEC project, funded by the European Union, for urgent decision making using ensemble simulation and real time analysis using in-situ techniques. This project demonstrates that early feedback of ensemble simulation has a meaningful impact on urgent decisions and could save time, resources and human lives. These use cases emphasize that scientific visualization tools for massive data can gather information from a real system, visualize it in 3D and send back information, actually closing the loop. Artificial Intelligence techniques are at the core of modern Digital Twins, because despite the data size growing, then the prediction, visualization and feedback loop should stay as close as possible to real-time. Deep Learning surrogate models are one of the techniques to ensure the expected performance, and the system should be able to run inference of Deep Learning models and use the output to display meaningful data. We will demonstrate that the in situ capabilities of scientific visualization tools could also be leveraged during the Deep Learning network training. These tools allow to visually monitor the output at the end of each epoch and efficiently tune the training process for better prediction and performance. In conclusion, Scientific Visualization tools are crucial for modern Digital Twins requiring advanced processing and prediction for large data to make decisions and influence the modeled system. They could be leveraged during the building phase and the tuning phase of the Digital Twin, in addition to being an essential part of the execution runtime of the Digital Twin.