From Excel Sheets to Self-Healing Networks: The Digital Twin Revolution in Telecommunications

By Nicolas Waern CEO, WINNIIO / Life Atlas | Co-Chair, Digital Twin Consortium (Telecom & Manufacturing)

The telecommunications industry sits at a paradox. It operates one of the most complex engineered systems humanity has ever built — billions of simultaneous radio connections across millions of cell sites, coordinated in real time — yet it still plans and optimizes that system with tools and methods that would be recognizable to an engineer from the 1990s.

I have spent over a decade working across digital twin implementations in smart buildings, manufacturing, healthcare, and telecommunications. The pattern is always the same: industries cling to spreadsheets and statistical approximations long after the complexity of their systems has outgrown those tools. Telecom is no exception. In fact, it may be the worst offender — because the consequences of suboptimal network planning are invisible to the people making decisions, buried inside aggregate KPIs that mask the physical reality of what is actually happening in the radio environment.

The shift from statistical models to physics-based digital twins is not an incremental improvement. It is a category change — from approximating reality to simulating it. And it is the single most important transformation the telecom industry will undergo in the next decade.

The Evolution of Network Planning

To understand where we are going, it helps to understand where we have been. The history of radio network planning is a story of increasingly sophisticated approximations — each generation better than the last, but none of them truly capturing the physics of radio propagation in a real environment.

1960s-2000s: Statistical Models — Circles on Maps

The foundation of radio network planning was laid in the 1960s and 1970s with empirical propagation models. Okumura's measurements in Tokyo (1968), later parameterized by Hata into the Okumura-Hata model, gave engineers a formula: plug in frequency, antenna height, distance, and environment type (urban, suburban, rural), and get an estimated path loss.

These models were revolutionary for their time. They allowed engineers to plan cellular networks with a calculator and a paper map. The COST-231 extension brought the models into the 1800 MHz range for GSM. Walfisch-Ikegami added street-level considerations.

But fundamentally, these were statistical averages. A cell site's coverage was a circle on a map — sometimes an ellipse if you were being generous about terrain. The models had standard deviations of 8-12 dB, which in practical terms means your predicted signal strength could be off by a factor of 6 to 16 in power. The models worked because networks were simple, traffic was voice, and margins were fat. You over-engineered everything and prayed the statistics held.

2000s-2015: Semi-Deterministic Tools — Better Approximations

As 3G and then 4G/LTE arrived, the industry adopted professional planning tools like Atoll, ASSET (now ASTER), and Planet. These tools introduced digital terrain models, clutter databases (land-use classifications that assigned propagation characteristics to categories like "dense urban," "forest," or "water"), and semi-deterministic ray-based approaches.

This was a genuine leap. Instead of circles, planners could generate coverage predictions that followed terrain contours and accounted for building density. Monte Carlo simulations distributed virtual users across the map to estimate capacity and interference.

But the clutter approach has a fundamental limitation: it treats all "dense urban" environments as equivalent. A 30-meter glass office tower and a 30-meter concrete parking garage are both "dense urban" — despite having radically different radio propagation characteristics. The models use average attenuation values for entire categories, smoothing over the physical reality that radio engineers deal with every day.

For 3G and early LTE deployments, this was sufficient. Coverage planning dominated over interference management. But as network density increased and spectrum became contested, the gap between the model's predictions and real-world performance began to matter.

2015-2025: Cloud-Native Platforms and Early AI

The O-RAN (Open Radio Access Network) movement, accelerated by Rakuten Mobile's greenfield deployment in Japan and championed by the O-RAN Alliance, introduced a fundamentally new idea: disaggregation. Separate the hardware from the software. Open the interfaces. Allow third-party intelligence — in the form of rApps and xApps — to optimize the network through standardized APIs.

Simultaneously, cloud-native network management platforms emerged. Operators moved from on-premises OSS/BSS stacks to cloud-based analytics. Machine learning entered the picture for anomaly detection, traffic forecasting, and mobility optimization.

But most of these early AI/ML deployments share a critical limitation: they operate on tabular performance management (PM) data — CSV exports of key performance indicators aggregated at 15-minute or 1-hour intervals. They know that handover failure rates increased between 2:00 PM and 3:00 PM on Cell A. They do not know why.

2025 and Beyond: Physics-Based Digital Twins

This is where the revolution begins. A physics-based digital twin of a telecommunications network is not a dashboard, not an analytics platform, and not a machine learning model trained on historical data.

It is a three-dimensional replica of the physical radio environment — buildings modeled as volumetric objects with material properties (glass, concrete, metal, wood), terrain at sub-meter resolution, vegetation, vehicles, even weather conditions — combined with the complete radio configuration of every cell site (antenna patterns, tilt, azimuth, power, frequency, neighbor relationships) and driven by Maxwell's equations through ray tracing or finite-difference methods.

This is not science fiction. The computational cost of full ray tracing has dropped by orders of magnitude over the past five years, driven by GPU acceleration (NVIDIA's Sionna and Aerial platforms), advances in 3D scene reconstruction (photogrammetry, LiDAR, even Gaussian splatting from commodity drones), and the maturation of voxel-based propagation engines that discretize the environment into volumetric elements.

The result: for the first time, we can simulate what the radio signal actually does in a specific location, not what a statistical model says it should do on average.

Why Raw Data Is Not Enough — The Context Problem

Here is a question I ask in nearly every engagement with a telecom operator or vendor: "You have a cell with an RSRP of -95 dBm at a specific location. Is that good or bad?"

The answer, invariably, is "it depends." And that dependency is precisely the problem.

RSRP (Reference Signal Received Power) tells you how strong the signal is. But signal strength without context is meaningless. What matters to the end user is signal quality — measured as SINR (Signal-to-Interference-plus-Noise Ratio). You can have a strong signal that is completely unusable because of interference from neighboring cells. You can have a weak signal that delivers excellent throughput because the interference environment is clean.

The relationship between RSRP and SINR is governed by physics: the positions and characteristics of every interfering cell, the reflections and diffractions in the environment, the antenna patterns, the beamforming configuration, the scheduled traffic load on each cell. None of this is captured in a CSV file of KPI counters.

This is the context problem. Operators collect terabytes of PM data — handover success rates, throughput measurements, RSRP/RSRQ samples, timing advance distributions, CQI histograms. They feed this data into AI/ML models. The models find correlations. But correlations without physical context produce optimizations that work in one cell and fail in the next, because the models have no understanding of why the numbers are what they are.

You cannot AI your way out of CSV data. A machine learning model trained exclusively on PM counters is doing sophisticated curve-fitting. It can detect patterns. It can predict trends. But it cannot reason about physics. It cannot tell you that the handover failure rate on Cell A is caused by a building that was constructed six months ago, creating a reflection path that causes the UE to see a strong signal from the wrong cell at the wrong time.

A digital twin can.

The Digital Twin as Simulation Laboratory

The most common misconception about digital twins in telecommunications is that they are dashboards — a visual representation of the network with real-time KPI overlays. That is a digital shadow: a read-only mirror of the current state.

A digital twin is a simulation laboratory. It is a physics-based replica that you can modify, experiment with, and stress-test without touching the production network.

Everyone in this industry is building a better library — better analytics, better dashboards, better reports, better visualizations of the same data. What we need is a laboratory. A place where you can ask "what if?" and get a physics-grounded answer before you make a change that affects millions of subscribers.

Train Before You Deploy

Consider the handover optimization use case. Today, operators tune handover parameters (CIO — Cell Individual Offset, hysteresis, time-to-trigger) reactively. A cell has high handover failure rates. An engineer adjusts the parameters. The change propagates. Sometimes it helps. Sometimes it shifts the problem to a neighboring cell. Sometimes it creates a ping-pong effect. The engineer adjusts again.

With a digital twin, you model the propagation environment at the problematic location. You identify the physical cause — a reflection, an overshooting cell, a coverage gap created by a new building. You simulate the effect of parameter changes before deploying them. You train an AI agent in the digital twin, where it can make thousands of adjustments and observe the simulated outcomes, before it touches a single production parameter.

This is not theoretical. NVIDIA's Aerial Research Cloud, combined with Sionna for link-level simulation, already enables this workflow. The O-RAN RIC architecture provides the deployment interface — rApps for near-real-time control, xApps for real-time optimization.

What-If Scenarios at Scale

The power of a simulation laboratory becomes exponential when you consider network-wide optimization. An operator planning a 5G densification strategy has thousands of candidate site locations, dozens of frequency band options, multiple antenna configurations, and complex inter-cell interference relationships.

Today, this planning is done with the semi-deterministic tools described earlier — or worse, with engineering judgment and spreadsheets. A digital twin allows you to simulate every candidate configuration, evaluate coverage, capacity, and interference for each scenario, and identify the optimal deployment plan before a single site is built.

The same applies to spectrum refarming, massive MIMO configuration, network sharing scenarios, and disaster recovery planning. Each of these is a complex, multi-variable optimization problem that is intractable with spreadsheets but straightforward in a simulation environment.

Smart Cities and Web4 — The Spatial Web

Telecommunications infrastructure does not exist in isolation. Cell towers are mounted on buildings, which sit on streets, which carry vehicles and pedestrians, which generate traffic patterns, which create demand that the network must serve. The physical context is everything.

This is where the digital twin revolution in telecom intersects with the broader spatial web — sometimes called Web4. The progression follows a clear trajectory:

Connect: Instrument the physical environment with sensors and connectivity. This is largely done.

Simulate: Build physics-based replicas that explain behavior, not just observe it. This is where the industry is now.

Automate: Use the simulation layer to drive autonomous optimization. TM Forum defines this as Level 3-4 network autonomy. Early deployments exist, but most operators are at Level 1-2 (assisted to partial automation).

Autonomize: Self-healing, self-optimizing networks that anticipate problems before they occur and resolve them without human intervention. This is the destination.

The spatial web framing matters because it reveals connections that siloed analytics miss. Customer churn, for example, is routinely analyzed as a CRM problem — pricing, customer service interactions, contract terms. But studies consistently show that network quality in the customer's home and workplace is one of the strongest predictors of churn. Map churn data onto the network topology, overlay it with the digital twin's coverage and quality predictions, and you can identify where network investment will directly reduce churn — a calculation that connects infrastructure spending to customer lifetime value in a way that spreadsheets never could.

Predictive maintenance offers another convergence point. A cell site digital twin that models not just the radio environment but the physical equipment — antenna systems, power supplies, cooling systems, fiber and microwave backhaul — can predict failures before they cause outages. Combined with weather simulation and traffic forecasting, the network can proactively reroute traffic around a site that is predicted to fail during a storm, before the first packet is dropped.

This is what self-healing means. Not a human looking at an alarm, opening a ticket, dispatching a technician. A digital twin detecting the preconditions for failure, simulating mitigation strategies, selecting the optimal response, and executing it — all before the subscriber notices.

The Business Case

The market for telecom network digital twins is growing rapidly. Industry analysts project the global Telecom Network Digital Twin market at approximately $2.7 billion in 2026, growing to $5.9 billion by 2031 at a compound annual growth rate of 16.6%. These numbers reflect both the network planning and optimization segments, as operators increasingly recognize that physics-based simulation is not a luxury but a necessity for 5G-Advanced and 6G network economics.

Several forces are driving this growth.

Network complexity has exceeded human cognitive capacity. A modern 5G network with massive MIMO, carrier aggregation, network slicing, and dynamic spectrum sharing has more tunable parameters per cell site than a human engineer can reason about. The industry built networks that require machine intelligence to optimize — but that machine intelligence needs a physics-based model to reason about, not just historical data.

O-RAN creates a market for third-party intelligence. In the traditional vendor model, Nokia, Ericsson, and Huawei each bundle their own optimization intelligence with their hardware — MantaRay SON, EIAP, and equivalent platforms. This intelligence is, by design, vendor-locked. It optimizes the vendor's own equipment and does not interoperate with competitors' gear. O-RAN disaggregation breaks this coupling and creates demand for vendor-agnostic intelligence platforms that can optimize a multi-vendor network. The opportunity for independent platform providers is substantial and growing.

The cost of not optimizing is increasing. Energy costs now represent 20-30% of operator OPEX, and the energy consumption of 5G base stations is significantly higher than 4G. Physics-based digital twins enable energy optimization — simulating the effect of putting cells to sleep during low-traffic periods, adjusting MIMO configurations, or tilting antennas to reduce coverage overlap — that can deliver measurable reductions in power consumption. When energy costs are $2-5 billion per year for a large operator, even a 5-10% reduction through better optimization justifies the investment in simulation infrastructure.

TM Forum Level 4 autonomy requires simulation. The TM Forum's Autonomous Networks framework defines Level 4 as "high autonomous" — the network predicts and prevents issues, optimizes continuously, and requires human involvement only for strategic decisions. No operator can reach Level 4 on analytics alone. You need a model of how the network should behave (the twin), a comparison of actual vs. expected behavior (anomaly detection), and a simulation of corrective actions (what-if analysis). The digital twin is the enabling infrastructure for network autonomy.

The Future is Physics-Based

The transition from statistical models to physics-based digital twins in telecommunications is not a question of if but when. The computational barriers have fallen. The 3D data acquisition costs have dropped by 90% in five years. The O-RAN architecture provides standardized interfaces for intelligence deployment. The market demand is clear and growing.

What remains is an execution challenge: building simulation platforms that are accurate enough to trust, fast enough to run in near-real-time, open enough to work across vendors and technologies, and practical enough that operators can integrate them into their existing workflows.

This is fundamentally a systems integration problem, not a research problem. The physics of radio propagation is well understood — Maxwell's equations have not changed. Ray tracing algorithms are mature. GPU compute is abundant. The 3D reconstruction pipeline from LiDAR and photogrammetry is production-ready. What the industry needs is not another breakthrough in any single component but the integration of these components into a coherent, deployable, vendor-agnostic platform.

The operators who adopt physics-based digital twins first will have a structural advantage: lower OPEX through optimized energy consumption, lower CAPEX through better site selection and configuration planning, higher revenue through improved customer experience and reduced churn, and faster time-to-market for new services and network slices.

The operators who wait will find themselves optimizing by spreadsheet in a world that has moved to simulation. And as network complexity continues to increase with 5G-Advanced, 6G preparatory work, and the densification required for fixed wireless access and industrial IoT, the gap between the leaders and the laggards will widen.

It was not a lack of stones that ended the Stone Age. It was the adoption of better tools. The same principle applies here. The Excel sheets are not going to get better. The networks are only getting more complex. The path forward is physics-based simulation — digital twins that do not just observe reality but replicate it, reason about it, and ultimately optimize it autonomously.

The future of telecommunications is not built on dashboards. It is built on laboratories.

Nicolas Waern is CEO of WINNIIO and Life Atlas, Co-Chair of the Digital Twin Consortium's Telecom and Manufacturing working groups, and Nokia Bell Labs Entrepreneur in Residence. He has advised over 100 organizations on digital twin strategy across smart buildings, healthcare, manufacturing, and telecommunications.