Building real-time digital twin instances with Function+Data Flow: user evaluation and extension for iterative pipelines
This paper presents an empirical user study demonstrating that the Function+Data Flow (FDF) visual language and its implementation in DesCartes Builder effectively enhance the accessibility and reliability of AI-based digital twin development for domain experts, while also motivating the proposal of a hierarchical extension (H-FDF) to support more complex, iterative pipelines.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the modern world, engineers and scientists often rely on "digital twins." These are not physical copies, but sophisticated software models that mirror a real-world object, like a bridge, a jet engine, or a power grid. The goal is to create a virtual version that learns and changes alongside the physical thing, allowing experts to predict failures or optimize performance before a problem ever occurs. To build these models, researchers often start with highly detailed computer simulations that are incredibly accurate but painfully slow to run. To make these models useful in real time, they must be simplified using machine learning, a branch of artificial intelligence that finds patterns in data. However, connecting these complex steps—turning slow simulations into fast, learning models—has traditionally been a messy, manual process. It requires deep programming skills and often results in fragile systems that are hard to fix or improve.
A team of researchers at Nanyang Technological University in Singapore and the CNRS in France set out to solve this engineering headache. They developed a new visual language called Function+Data Flow, or FDF, designed to let experts build these complex machine learning pipelines by dragging and dropping blocks on a screen, rather than writing lines of code. They built a tool called DesCartes Builder to put this idea into practice. To see if this approach actually worked, they invited twenty-five researchers and students to a workshop. These participants were asked to build a real-time digital twin prototype that could predict how much a material would stretch and deform under stress, a task critical for structural safety. The participants used the new tool to assemble the necessary steps: taking raw data, simplifying it, training a machine learning model, and testing the results.
The results of this experiment were encouraging. The participants found the tool surprisingly easy to use, even though it was still in an early stage of development. On a standard scale used to measure how user-friendly software is, the tool scored well above the threshold for "unacceptable," with a median score of 72.5 out of 100. This suggests that the visual approach successfully lowered the barrier to entry, allowing domain experts to focus on the engineering problem rather than the coding details. The study also revealed that the tool was particularly effective for those with technical backgrounds, while those less familiar with the specific concepts faced a steeper learning curve. The researchers noted that the visual nature of the tool made it clear how data moved through the system, and participants appreciated the ability to swap different machine learning algorithms with a single click, a task that would normally require rewriting large sections of code.
However, the study also highlighted where the system needs to grow. Some users felt the tool lacked enough built-in guidance, such as pop-up hints or tutorials, to help them understand the purpose of each block. Others found the error messages confusing when things went wrong, wishing for clearer explanations of what caused a failure. To address these limitations and the need for more complex workflows, the researchers proposed an extension to their system called H-FDF. This new version introduces a way to create loops within the design, allowing the system to repeat a process until it reaches a specific goal, such as a model becoming accurate enough. This is a significant step forward because it allows for more advanced techniques, like training two models together to improve each other, without making the overall design chaotic.
The researchers compared their new approach to existing tools used in the industry, such as KNIME, which is a popular platform for data analysis. They found that while those tools can handle loops, the way they do so often mixes up the flow of data with the flow of control, making the diagrams cluttered and hard to modify. In contrast, the new H-FDF approach keeps the data flow and the looping logic separate and clearly defined, making it easier to change how the system behaves. By testing a complex scenario where two models were trained simultaneously to learn a dynamic system, the team showed that their new method could handle these intricate tasks more cleanly. The study concludes that while the current tool is a promising start, turning digital twin engineering into a disciplined, accessible practice will require continued refinement, better documentation, and the full implementation of these new looping capabilities. The work suggests that with the right visual tools, the complex art of building digital twins can become a more reliable and accessible discipline for the experts who need them most.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.