Radio access network

(RAN) digital twins

Why are they essential to optimizing
RAN performance
2 Ericsson | Radio access network (RAN) digital twins

The digital twin era sparks

a RAN revolution
The digital twin era At the heart of the digital twin lies
a fascinating principle: computational
sparks a RAN revolution irreducibility. This concept is invaluable Digital twin
Radio access network (RAN) optimization for understanding the intricate dynamics Digital twins are dynamic
has greatly advanced since the advent of of today’s complex systems, such as virtual replicas of physical
“digital twins”. By creating simulated virtual live RANs. objects, processes and
testing environments, digital twins let us In his influential book, “A New Kind of
Science”, Stephen Wolfram introduces the
systems that enable
safely run “what. if” scenarios and analysis
concept of computational irreducibility, and risk-free testing and
in risk-controlled RAN environments.
These sophisticated, cloud-native tools its profound implications for understanding prototyping, improved
harness real-world data to enable simulations and simulating complex systems. Today, efficiency, productivity,
of the intricate behavior of particular RANs. industry experts are deftly employing and safety, as well as
In doing so, digital twins offer unparalleled this concept to construct digital twins of
RANs, thereby unleashing the power of accelerated time-to-market.
insight into network performance and
optimization. Such a breakthrough approach computational irreducibility to advance
Source: Ericsson
marks a significant stride in our relentless our pursuit of network performance
pursuit of network spectrum efficiency and optimization.
user experience enhancement.

Computational irreducibility
Before we dive into the implications for RAN simulations, let’s first understand what computational irreducibility means. In its
essence, this principle suggests that certain systems or processes are so complex that there is no way to predict the outcome
without emulating the most systemic operations in an explicit manner. For these systems, there is no “shortcut” or single,
closed-form formula that can accurately estimate the future state without going through all preceding intermediate states.

By contrast, in a computationally reducible system, you can predict the outcome using a simpler or more efficient computation.
An example would be a perfectly spherical ball rolling down a frictionless incline. Newton’s laws of motion allow you to
calculate the ball’s future position without having to simulate every detail of the ball’s movement.
3 Ericsson | Radio access network (RAN) digital twins

Applying computational
irreducibility to RAN
digital twins

As we attempt to simulate RAN Initially, we run a measurement and authentic partial replica of the RAN
performance, we confront a task that campaign for a specific RAN cluster— cluster. This incorporates radio frequency
seems to be computationally irreducible. the optimization target. We ingest live (RF) measurements, UE traffic, and mobility
Working with CSP’s around the world such network data, down to individual User patterns for that measurement period.
as Swisscom, we’ve successfully executed Equipment (UE) granularity, to construct This first step alone has inherent value.
several deliveries in CSP’s live networks, a static digital twin (DT). This static DT For instance, it can help the operations
targeting large-scale network optimization. corresponds to a timespan indexed with team visualize network performance.
Here’s how the system works: Recording Output Periods (ROPs), usually More importantly, it lays the foundation
set at 15-minute intervals. The goal of this for downstream tasks, such as running
step is twofold. First, we create a virtual “what-if” analysis.
4 Ericsson | Radio access network (RAN) digital twins

Real network A

RAN digital twin in a nutshell

Data inputs: B
De-identified live network data at user
equipment granularity and network
parametrization – Configuration
Management (CM)
Time domain: De-identification
Recording Output Periods (ROPs) layer
Data operations:
Analytical and/or Machine Learning Static DT
Behavioral Models (BMs) to emulate network
performance given CM parameter changes
Data outputs:
Synthetic Performance Management (PM) Static DT
counters timespan
ROP_96

ROP_95

...
ROP_1

Figure 1: Static digital twin generation with a live network measurement campaign

On top of the static DT, the procedure Operating on a modular and interactive flexibility also allows for the seamless
instantiates a dynamic digital twin that framework, our digital twin facilitates the integration of various optimization
considers Configuration Management (CM) implementation of a wide array of use algorithms to suit specific use cases, such
parameter changes of cells in the cluster case scenarios. For instance, we utilize as incorporating mathematical or even
as inputs. Based on this, it subsequently reinforcement learning (RL) to construct an rules-based algorithms when appropriate.
triggers a series of behavioral models that AI optimizer. The synthetic PM counters are An added benefit is the capacity for human
simulate interactions between the UEs and used to construct state and reward for an interaction within this modular design,
the radio base stations (BSs), and produces RL agent, and that RL agent’s action output fostering hands-on learning and training
corresponding synthetic Performance then becomes the input to the next dynamic opportunities.
Management (PM) counters for each cell DT inference. This framework’s inherent
in the cluster.
5 Ericsson | Radio access network (RAN) digital twins

In figure 2, the computation process of values, the only way to get the outcome
a single dynamic digital twin inference is to emulate it. This is a clear example of
is visualized with data represented by computational irreducibility in action. Even
grey boxes, and computation denoted when we understand the rules governing
by arrows, such as behavioral models the system (i.e., the behavior models),
(BM). Since this complex system cannot we can’t shortcut the computation to
be estimated using a formula with new predict the system’s state given new CM
Configuration Management (CM) parameter parameter changes.

Figure 2: Dynamic RAN digital twin in action as it receives CM parameter changes

proposed by an AI optimizer, and outputs new values for cell-level synthetic counters.

AI optimizer

Dymamic DT
(Initialize with static DT)

BM_1 BM_2 BM_3 (...)

CM parameter Data
changes operations

Static DT
timespan

ROP(N)

N N N
...

...

...
ROP(N-1)
N-1 N-1 N-1
...
ROP(1) 1 1 1

Coverage Interference Energy consumption Behavior model

(...)
counters counters counters execution
RAN digital twin in a nutshell
Data inputs:
De-identified live network data at user equipment granularity and network parametrization – Configuration Management (CM)
Time domain:
Recording Output Periods (ROPs)
Data operations:
Analytical and/or Machine Learning Behavioral Models (BMs) to emulate network performance given CM parameter changes
Data outputs:
Synthetic Performance Management (PM) counters
6 Ericsson | Radio access network (RAN) digital twins

The role of machine

learning in applying
digital twins to RANs
In a system as complex as a RAN, covering
every possible state might not be feasible.
• Overfitting: ML models can overfit
the training data, resulting in poor
performance when faced with unseen
data or states.
• Complexity: Since the complexity of the
ML model lifecycle management process
might be comparable to the original
system’s complexity, this potentially offsets
any computational “shortcut” benefits.
While we appreciate the potential of ML
models to enhance our digital twin, we must
not lose sight of the critical importance of
generalizing unseen data or states. This is
especially vital due to the cascading effects
of two types of prediction errors. First, a
prediction error from one behavior model
can propagate to the next within a single
dynamic digital twin inference, potentially
skewing the results.
More importantly, as our AI optimizer
iteratively interacts with the dynamic
digital twin, the more exploration, or
optimization iterations, it performs. And
the further the state of the digital twin
deviates from its original baseline. Despite
these iterations, maintaining the accuracy
In building digital twins for RANs, our model that accurately predicts
and trustworthiness of prediction results
behavioral models must be firmly grounded unobserved states can be difficult,
remains a challenging but crucial task.
in analytical principles. These principles mainly because of inherent limitations or
Here, our behavior models’ generalization
often mirror the laws of physics and system complexities within the system. In such
capabilities are indispensable to ensuring
rules, which generalize well even to unseen situations, a viable approach is to gather
the authenticity of our digital twin, even as it
states. Deep-seated knowledge of RAN a training dataset that incorporates
evolves away from its real-world baseline.
accumulated over the years forms the ground truth information, which can often
In response to these challenges—and
foundation for these models, providing be costly to collect. We then employ ML
to manage the ML models’ limitations—
reliable and interpretable simulations that techniques to approximate the unknown
we primarily rely on analytical models.
emulate network behavior. function, linking input features and output
Depending upon the focus of a specific
Machine Learning (ML) models can ground truth labels. During the inference
analysis, we decide which aspects to explicitly
also serve as complementary tools in phase, the pre-trained ML model is then
“emulate”, and what elements to encapsulate
certain scenarios within our digital twin. utilized to predict labels that cannot be
within an ML model. This approach gives
We carefully incorporate these ML models, directly observed within the system.
our architecture flexibility and modularity,
employing them strategically alongside • Speeding up inference: Once trained, empowering us to delve deeper into a
our primary analytical behavior models. ML models can provide rapid predictions, particular problem, while approximating less
The result is a thoughtfully designed which are advantagous to reducing time critical mechanisms for that analysis.
hybrid system that blends the reliability of and the consumption of computational Our profound domain knowledge
analytics-based models with the versatility resources. In other words, ML models also acts as an inductive bias in the ML
of selective ML-based behavior models. serve as computation “shortcuts”. model development process, guiding us in
ML models can be beneficial in two
Despite these potential benefits, it’s crucial selecting the algorithm and structure of the
key ways:
to recognize the limitations of ML models: ML model. As a result, we can manage the
• Approximate complex functions:
• Data requirements: ML models need challenges of generalization and complexity
Constructing an analytical behavior
vast amounts of diverse training data. more effectively.
7 Ericsson | Radio access network (RAN) digital twins

Distinguishing Ericsson
RAN digital twin from
conventional simulators

With respect to RAN digital twins, there are

Ericsson RAN Standard static Standard dynamic
two conventional approaches, static and
digital twin simulator simulator
dynamic simulators, and the enhanced
Ericson’s solution. The static and dynamic Small, directly operate Large*, operate on Large*, operate on
Reality gap
approaches are considered de facto industry on real-world data simulated data simulated data
standards. Without any internal history of past
output and input values, a static simulator Scalability to
Good Good Not scalable
relies entirely on a function (f), which is based large network
solely on current inputs (u). The output of the Adaptability to
system, in this case, is solely dependent on Good Not adaptable Not adaptable
network evolution
these inputs.
Full functionality
Functionality Limited functionality** More functionality
is possible

Accurate predictions to
Authenticity Captures basic trends
a specific RAN cluster
* The gap can be minimized by statistically modeling UE traffic patterns using real-world data, which requires additional effort
**Has the potential to support wider functionalities
8 Ericsson | Radio access network (RAN) digital twins

On the other hand, a dynamic simulator standardization and feature development predicting the system performance of a
maintains an internal history of prior inputs, activities. And they can effectively evaluate specific scenario. In particular, these
internal variables, and scenarios. This generic scenarios and offer valuable insights simulators often fall short due to the
enables it to model the varying behavior of into overall system behavior. complexity of accurately defining all inputs,
a system at various times or under different Despite their capabilities, both types of such as spatial traffic distribution, mobility
conditions. It’s important to acknowledge simulators share a common limitation: they patterns, traffic patterns, and exact
that, both static and dynamic simulators struggle to accurately replicate the propagation mechanisms. This limitation
have their respective strengths, and—in the real-world RF (radio frequency) highlights the appeal of a purely data-
wireless communications industry, they have environment and traffic profile of the driven approach.
been used for various research, operations, network. This is particularly evident when

So, what distinguishes the Ericsson RAN digital twin from industry-standard simulators?
Our RAN digital twin brings a unique feature set to the table, including the following:

Real-world data
Unlike traditional simulators that often base UE patterns on widespread theoretical assumptions, our digital
twin directly operates on actual live network data. Each data frame in our system is a joint RF footprint of a
real-world measurement, coupled with UE data traffic consumption information. Moreover, our model complies
with stringent privacy protection measures, such as GDPR regulations. For example, our solution works with
anonymized data, where UE identities are replaced with dummy IDs, making it impossible to track individual
information. This approach ensures we maintain the UE-level visibility necessary for accurate digital twin
modeling, while fully respecting privacy concerns. By responsively handling granular, authentic data, we
establish a realistic basis for simulating a specific RAN cluster.

Scalability and adaptability

Another key distinguishing factor of our digital twin is its scalability and adaptability. Our system is engineered
to model large RAN clusters, handling up to thousands of 4G/5G cells along with their connected UEs over a
certain time period. For example, a typical measurement campaign usually takes a few hours per day over a
few days to make sure the data is sufficiently representative to capture the real-world RF environment and its
UE traffic patterns. While this degree of scale is vital to solving network optimization problems, it’s typically a
challenge for many other traditional simulators. Furthermore, as networks evolve and expand, our digital twin
accommodates these changes by ingesting new data streams. This adaptability ensures that our tool remains
relevant to its real-world target, regardless of how the network landscape changes.

Functionality
Compared with traditional simulators, our current digital twin implementation supports a much smaller subset
of tunable CM parameters. When compared to real base stations, our RAN digital twin generates a limited
number of synthetic PM counters, focusing on those we believe are essential to the optimization agent. These
choices were strategically made to prioritize realism and practicality within given resource constraints. In
contrast, traditional simulators typically offer support for a broader range of CM parameters and incorporate
more complex behavior models, especially in the Radio Resource Management (RRM) layer. While developing
our RAN digital twin, we carefully simplified and abstracted behavior models to ensure equivalence. And, while
expanding behavior models for additional CM parameters is a complex and resource-intensive undertaking, we
plan to further enhance the capabilities over time.

Acknowledging the significance and potential of other simulators within the industry is essential. Rather
than replace these simulators, our goal is to introduce an alternate, data-driven approach, precisely
designed for large-scale network optimization use cases.
9 Ericsson | Radio access network (RAN) digital twins

In conclusion
Computational irreducibility underscores the deep complexity of systems like RAN.
It challenges us to develop detailed simulations, such as our RAN digital twin, rather
than seeking shortcuts that don’t exist.
While ML models serve as valuable complements to our primary analytical
models, we remain aware of their limitations, including extensive data needs and
potential overfitting. Leveraging our deep domain knowledge, we apply strategic
methodologies in the design of ML models to maximize the robustness, performance,
and cost efficiency of our solutions.
Based on our live network experiences with tier 1 communications service providers,
our RAN digital twin outperforms traditional simulators when dealing with network
impacting parameters by ingesting live network data and modeling large RAN
clusters. While it can’t model every aspect of a RAN due to computational and data
constraints, we’re tackling the limitations and complexity, and striving to realize an
optimal solution which maximises the potential to model dynamic RAN networks.
Computational irreducibility is a testament to this complexity, keeping us humble
and inspiring continued innovation in our quest to make the best-performing
networks even better.
10 Ericsson | Radio access network (RAN) digital twins

Further reading
Network digital twins – outlook and opportunities - Ericsson

Ericsson cognitive technologies

Please visit us to find out more about Ericsson Cognitive Software offerings
Boost network performance with Cognitive Software - Ericsson

Request demo
Demos are available upon request, please Contact us
11 Ericsson | Breaking down OSS/BSS silos for better 5G monetization

About Ericsson Ericsson enables communications service providers to

capture the full value of connectivity. The company’s
portfolio spans Networks, Digital Services, Managed
Services, and Emerging Business and is designed to help
our customers go digital, increase efficiency and find new
revenue streams. Ericsson’s investments in innovation
have delivered the benefits of telephony and mobile
broadband to billions of people around the world. The
Ericsson stock is listed on Nasdaq Stockholm and on
Nasdaq New York.

www.ericsson.com

Ericsson.com The content of this document is subject 28423-FGB1011026 Uen

to revision without notice due to © Ericsson AB 2023
continued progress in methodology,
design and manufacturing. Ericsson
shall have no liability for any error or
damage of any kind resulting from the
use of this document.