Glasgow digital twin tests networks 25,000 times faster

Researchers at the University of Glasgow have developed a machine learning-based digital twin system for testing computer networks. In trials, it tested networks 25,000 times faster than a traditional simulator.

The work was led by Shenjia Ding, a research student in the university's School of Computing Science. Ding used automatically generated digital twins to test two complex computer networks in the US and Europe, with 12 and 37 nodes respectively.

In the trial, the digital twin took 4.78 seconds to test network speed, compared with 33 hours for a conventional simulator used as a benchmark. The model maintained accuracy while sharply reducing the time needed to run the tests.

Traditional network testing relies on simulators to recreate real-world conditions so engineers can assess performance, security and reliability. The process can be slow and often requires specialist expertise to configure and interpret the models.

The Glasgow team used Automated Machine Learning, or AutoML, to build the digital twin. This approach automates parts of machine learning model development and can lower the barrier for users with limited experience in the field.

The researchers tested the system under six types of traffic, including web browsing, video streaming and file downloads. They also introduced continuous congestion and background noise to mirror conditions seen in operational networks.

Digital twins are virtual replicas of physical systems or processes that can be used to test changes before they are applied in the real world. In network management, they offer a way to examine how systems respond to shifting traffic patterns or faults without disrupting live services.

"Our results show that testing computer networks with automatically generated digital twins can achieve high accuracy and significantly faster speeds than traditional simulator-based testing," said Shenjia Ding, a research student at the University of Glasgow.

"We're demonstrating a very promising alternative to manual, time-consuming testing that also relies heavily on professional expertise."

The study suggests digital twins could become more useful as internet traffic and data volumes continue to increase. Faster test cycles may help researchers and network operators assess more scenarios in less time, especially as systems grow more complex.

Beyond telecoms

Paul Harvey, a co-author of the research and senior lecturer in the School of Computing Science, said the same methods could extend beyond computing networks. He pointed to transport systems as another area where rising data volumes and network pressure are becoming harder to manage.

"Transport, like computing, is seeing enormous growth in data volumes, and in both instances the pressure on the communications networks carrying all this data is immense," Harvey said.

"By proving that we can use machine learning to build digital twins, which is another time-consuming and laborious task, we are highlighting the huge potential of this research to test and optimise transport and other networks that we rely on daily."

Harvey is also a co-investigator for TransiT, a UK research hub focused on using digital twins and related tools to support transport decarbonisation. The project is a collaboration between eight universities and nearly 70 industry partners, jointly led by Heriot-Watt University and the University of Glasgow.

The latest work could support TransiT's aim of creating a digital twin factory designed to automate the production of digital twins for transport settings. That aligns with a broader push across industry and infrastructure to build digital models more quickly and at lower cost.

The paper is titled Automated Digital Twin Generation for Network Testing: A Multi-Topology Validation. Its co-authors are Paul Harvey and David Flynn at the University of Glasgow.

The next stage of the work will examine how the digital twin can be updated over time, what it costs to maintain, how it performs in real-time network environments and how it compares across a wider range of network scenarios.

For telecoms researchers and operators, the central finding is the scale of the time reduction shown in the tests: 4.78 seconds for the digital twin, compared with 33 hours for the traditional simulator.