The future of digital twins: what will they mean for mobile networks?

The idea of duplicating one’s mind into a machine, decoupled from the physical body has long fascinated humans. Now it’s reality – welcome to the age of digital twins. Maybe not for humans (yet), but for the systems around us. Let’s get to know digital twins and what they mean for optimizing and automating mobile networks.

In the 1930 Science Wonder Stories, featuring ‘The Infinite Brain’, inventor Anton Des Roubles uploads his entire brain into 200,000 memory cells and continues to live in machine-form after death, decoupled from his biological existence. That story quickly takes a turn for the worse, so we’ll leave Anton and his destiny there for now, with the observation that today we count memory in gigabytes and terabytes, but holding an entire human mind in a device is still very much science fiction.

There are many stories on the theme of artificial brains and cloning in popular culture and science fiction, but a more technological precursor of today’s digital twins can be found in the 1992 book ‘Mirror Worlds: or the Day Software Puts the Universe in a Shoebox...How It Will Happen and What It Will Mean’. Here, the computer science professor David Gelernter outlines a future where computer systems are interconnected globally and observe in real-time the physical world around us. The resulting images and representations can be presented to humans through a pane of glass, accurately mirroring the real world. In his vision, we can also interact with the presented images, controlling things in the real world through this mirror.

These were still only ideas, and the first real use of the digital twin concept was done in 2002, when Michael Grieves applied it to product lifecycle management in manufacturing. Since then, the concept has evolved and been applied to many areas. Here we’ll start with an overview of what digital twins are and their typical use cases, then move on to see how these ideas can be used in optimizing and automating our networks.

What is a digital twin?

In essence, digital twins are software representations of assets and processes, which are enhanced with capabilities not present in the real-world entity. There are many proposed definitions which emphasize different aspects of digital twins, and the concept is still evolving, but there are typically a few common characteristics listed here and illustrated in the picture below:

• First, you need data models and data structures to represent the observations, state and relations of the real-world objects of interest.
• Second, you need to populate these models with knowledge, for example, actual data from the specific real-world objects to create the actual digital twin instance. In many cases this data collection is done continuously to enable an accurate and up-to-date view in the twin.
• Third, you need tools that operate on the data to add value. These are the key to unlocking the capabilities and benefits of the digital twin. Arguably the most integral group of tools are the analytics models to extract insights from the digital twin, spanning from simple data retrieval to complex algorithms used to predict future behavior, simulate different scenarios and other analytics tasks. But there are also tools for actuation to add improved steering capabilities to the real-world objects.
• Finally, you need ways to interact with the twin through different APIs & visualization The insights you gain will be used to make better decisions in the real-world, either via a human or by direct actuation from the digital twin itself.

Conceptually this looks simple, but often advanced software is needed to realize these ideas, as we will see later. Data management will be a challenge as data volume grows and execution of simulations needed for analysis can be very time consuming.

Figure 1: The structure of a digital twin.

How are digital twins used?

Digital twins are often applied to big expensive machines, such as jet engines or power plants. In these cases, there are high costs associated with unexpected downtime and the consequence of a failure can be severe. With the enhanced knowledge you get from the digital twin, it is possible to optimize service windows and predict when different components need to be replaced. The advances in compute capacity and recent progress in AI-based analytics makes the concept attractive to a wider set of use cases and in the IoT area, ranging from simple devices like temperature sensors and light bulbs to entire cities with a complex web of traffic, utilities and buildings. There are many different initiatives pushing this ahead, both in standards and commercial offerings.

Turning our attention back to mobile networks, they certainly fit the criteria of big machines with high consequences for downtime and failures, as well as the added complexity of massive geographical scale. With this in mind, it does make sense to apply digital twin concepts here as well, to add value in both the operations phase and the development phase. In fact, in some ways existing solutions already use core ideas from digital twins. Network planning tools have long been used to understand the current network situation and plan for upgrades to satisfy future demands. Typically, these are not real time, but they use the same basic concept of data collection, analysis and prediction to support decision processes for network build-out to satisfy future demand.

There are also solutions to optimize site management and field operations, such as Ericsson’s Intelligent Site Engineering, which uses photos to create virtual 3D models of radio sites. Site deployment and maintenance can then be planned in detail from the office, reducing truck rolls, site visits and tower climbs. The Advanced Microwave Insights is a recent addition which analyzes data from microwave networks in near real time, to improve performance and save costs.

These examples are for quite specific use cases and we continue to identify additional use cases and to see how digital twins could be integrated in our 6G networks system to support development and operations. Some digital twin applications would be quite specific, focusing on one part of the network like a site or a radio cell at a high level of detail. On the other extreme, we have twins with an end-to-end view, leading to a higher-level analysis of services or network infrastructure. Other twins could focus on a specific domain like the RAN or edge clouds.

Figure 2: Areas of interest for digital twins in mobile networks.

What applications do we see for 6G networks?

Let’s take a closer look at some areas we have identified where network digital twins have interesting use cases.

New services - In 6G we expect a much wider variation of services to be deployed compared to today. There would be specialized services for small groups of users like emergency personnel or self-assembling robots. For the mass market, immersive sports events would be offered with enhanced interaction and augmented or virtual reality tailored to the specific occasion. Such services have high requirements on performance and their related service-level agreements (SLAs) have tough KPIs to fulfil. Using a network digital twin, we create a replica of the current network state (the “twin”), then deploy the new service in the twin, using analytics models to simulate and analyze the service performance and its impact on the other, already existing services. If everything looks good, we can go ahead and launch the services. But if there are potential issues, we can use this new insight to adjust the service parameters or prepare the network for the increased load. This could be both updated usage quotas or actual network upgrades to support the increased load.

Future scenarios and deployments – In making network upgrades and build-out, it is important to understand when and how to do this in the optimal way. The service requirements are projected by analyzing current trends and adding expected future developments. Once the needed network and compute capacity has been identified, a digital twin could be used to test different future scenarios. Options for coverage improvement and extension of transport and cloud infrastructure could be analyzed in greater detail and the resulting performance compared. Another separate but related aspect is the impact on services from different failure cases. In a digital twin, an operator would be able to identify weak spots and quantify consequences of failures. In doing so, it is possible to evaluate in greater detail both fulfilment of customer expectations and financial risks from SLA breaches. These insights can then be taken into account when planning for the future, to optimize the investments and funnel them to where they have the best economic value.

Training of AI – Most AI algorithms need large amounts of realistic data to be trained. A digital twin could be used to train algorithms when the real network does not provide enough data or, as in the case of reinforcement learning, it is risky to apply an AI algorithm under development to the real network. In a digital twin, one can also create different problem situations and failures which are rare in network operations, enabling the AI algorithm to train for such situations, to recognize and act correctly upon them in cases they do occur in live networks.

Configuration and function changes – When you apply a new configuration, a new AI model or a new software in a network it is important to know how it performs before introducing it to the whole network. In today’s continuous integration and deployment (CI/CD) pipelines, canary testing is done to validate that services perform and function as expected. A digital twin would be a further tool to test impact in a safe environment, for example, by evaluating the planned steps for deploying the new software in the cloud environment to make sure that the resources available are sufficient to do the test and, if successful, switch over to this new version.

In autonomous networks, a network management system takes actions on its own to make sure services have the expected performance and that the objectives defined for network operations are fulfilled. Understanding the impact of these actions is essential and a digital twin could be used to test different actions, evaluate their impact on relevant KPIs and decide on the best options before they are implemented for live services.

Figure 3: The process view of a digital twin.

Click to view an enlarged image

What are the challenges with realizing network digital twins?

So, how can we actually realize this? There are a few different parts which need to be in place. A key enabler here is data. Data-driven network architecture is crucial for developing the infrastructure required for digital twins – we need to know the state of the network in real time, or at least near to real time. Measurements need to be done at a greater level of detail than what is typically done today, gathered in databases, and made available for analysis. This is an area where there are currently significant efforts ongoing, both to capture data and to gather this data for later use. For some use cases, data availability is already sufficient to create digital twins, whereas other, more advanced use cases still need improved capabilities in this area. Typically, more detailed analyses would require more detailed data, and one would need to balance the cost of gathering and managing data with the gains achieved for different use cases.

In a digital twin, both historical and current data will be used to create models to understand behaviors, analyze the current situation and make predictions about the future. There is a trade-off here also, between complexity, accuracy and the computational resources needed to run the models. To some extent, we are helped by computer hardware improvements and availability of cloud environments. But even today, detailed models can in practice only be executed for quite small systems and few users. In many cases, there is also a time aspect – a need to evaluate models much more quickly in the twin than in the real network. Otherwise, we would get an answer, but it would be too late for it to be useful. To fulfil these conflicting requirements, there would be a mix of different models depending on the use case at hand, the network domain of interest, and the level of detail needed. Some would be based on different behavioral and physical models whereas others would use AI and be entirely learned from data.

To complete the digital twin, various tools and APIs for interaction need to be developed. Humans will create different scenarios and test configurations and control their execution in the twin. New visualization interfaces will also be required to present results and findings in a quick and accessible way, supporting timely decision processes. In a similar way, the digital twin must also be directly accessible from software used in development and operations, without direct involvement of humans. This access would be through management systems and AI training pipelines, but also other digital twins. In such a way, a network digital twin could be part of a larger digital twin ecosystem which could include, for example, city planning, transportation, smart buildings, and power grids.

The next steps – the evolution of digital twins

With the outlook described above and the opportunities identified, how can we realize the desired capabilities in this area? There are two tracks we need to evolve simultaneously to gradually introduce new solutions into network operations and development.

On one hand, specialized solutions will continue to develop for the most attractive use cases and new capabilities are already being added to existing solutions. As these evolve and become more advanced, similar to the evolution we see in the IoT area, common components and frameworks will likely emerge and gradually evolve into generally applicable solutions and a platform for new digital twins.

One the other hand, there is a need to understand the bigger picture on how digital twins will fit in an evolved network architecture. There are currently efforts ongoing in standardization (IETF & ITU-T) to define architectures for digital twins in a network context. Once these mature, they will form the basis for creating frameworks and solutions for different use cases, both those mentioned above and future use cases yet to be discovered.

To quote the prophetic words printed over 90 years ago in Science Wonder Stories: “And lest you jump to the conclusion that intelligent, or quasi-intelligent, machines are pure figments of the imagination, remember that already, today, we have machines that can “think" faster and better than any human being… There is no question that, in the future, even more wonderful machines will be evolved along these lines.”

About the Author

Ericsson is one of the leading providers of Information and Communication Technology (ICT) to service providers. Ericsson enables the full value of connectivity by creating game-changing technology and services that are easy to use, adopt, and scale, making our customers successful in a fully connected world.

Lars Magnus Ericsson founded Ericsson 145 years ago on the premise that access to communications is a basic human need. Since then Ericsson has continued to deliver ground-breaking solutions and innovate technology for good.