AI-Based Radio Propagation Modelling for Closed-Loop Autonomous RAN Optimisation

Autonomous RAN optimisation lacks propagation awareness — but AI holds the key to the solution. For RF and RAN engineers, the compromise between accuracy and computational speed in radio propagation modelling has shaped many decisions in network planning. Forsk has developed the AI technology that eliminates the need to compromise, and the implications for autonomous network optimisation are significant.

The trade-off every RF engineer knows

Ray-tracing models produce path loss predictions of high fidelity. By simulating electromagnetic propagation — accounting for reflection, diffraction, penetration, indoor/outdoor transitions, vegetation, and geoclimatic effects — they generate a detailed picture of how radio signals behave in a given environment. That level of precision is exactly what engineering-grade planning and optimisation requires.

The cost is computation time. For complex propagation environments, ray-tracing is expensive to run. Historically, that cost was acceptable. Network planning operated on cycles measured in days or weeks, and predictions ran in batch. Speed was a convenience, not a hard requirement.

Statistical and empirical models occupy the other end of the spectrum: faster, but less precise. They are useful for rapid iteration and large-scale coverage estimation, but they sacrifice the accuracy that makes predictions reliable for engineering decisions in dense or heterogeneous environments.

For static planning workflows, this trade-off was manageable. Engineers selected the model appropriate for the task at hand. The constraint only becomes structurally problematic when the timescale of the task changes.

Why autonomous networks change the requirement

Closed-loop autonomous network optimisation requires near real-time insight into radio conditions. An autonomous system that can detect coverage degradation, evaluate corrective actions, and apply changes to live network elements needs to understand how radio waves are propagating — not as a batch output from the previous night's run, but at operational timescales.

Today's self-optimising network (SON) capabilities have made meaningful progress on parameter automation. They adjust power levels, handover thresholds, and load balancing based on performance metrics. What they do not do is incorporate propagation modelling into the optimisation loop. Decisions are made without a continuously updated model of the radio environment.

The consequence is a structural gap: closed-loop automation that is genuinely propagation-aware requires path loss predictions fast enough to feed a decision loop. At the speeds closed-loop systems need, ray-tracing is too slow. At the accuracy engineering decisions require, statistical models are insufficient.

Bridging that gap is what AI-based radio propagation modelling makes possible.

An AI model trained on ray-tracing outputs

Forsk has developed an AI propagation model built on a convolutional neural network (CNN) trained on the outputs of Aster, Forsk's calibrated ray-tracing propagation model. The CNN learns the relationship between geospatial inputs — terrain data, clutter classification, building geometry — and path loss, and applies that learned mapping at inference speed.

The architecture is deliberate. Because the model is trained on calibrated Aster predictions, it inherits Aster's accuracy characteristics. Because it runs inference rather than full simulation, it operates at a fundamentally different speed. The result is a model that does not force engineers to choose between precision and performance.

This capability is now a GPU-accelerated compute component of Naos, Forsk's cloud-native platform for RAN planning and optimisation 

What the performance figures show

On NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, Forsk's AI propagation model runs up to 200 times faster than the Aster ray-tracing model it was trained on — while achieving less than 1 dB mean absolute error.

That combination is what makes near real-time propagation modelling operationally viable. A full-resolution path loss prediction can now be generated within a timeframe relevant to autonomous network decision-making, rather than as an output of an hours-long batch process.

The white paper documents the full benchmark analysis across NVIDIA GPU generations, alongside direct accuracy comparisons against CW measurement data and calibrated Aster model outputs. Those results are where the engineering case is made in detail.

Use cases this enables

Speed and accuracy at this scale unlock use cases that have so far been impractical at operational timescales.

Network self-healing — the automatic detection and correction of coverage degradation caused by antenna faults, interference, or environmental change — requires the network to understand, in near real time, how a parameter change affects coverage across a cell. That understanding depends on propagation modelling.

Automated antenna tilt optimisation follows the same logic. Adjusting electrical tilt to balance load or restore coverage is a well-understood engineering task. Doing it reliably and autonomously — without human review of each adjustment — requires a propagation-aware model fast enough to evaluate the impact of changes before they are applied to live network elements.

These are not speculative applications. They are engineering problems with known mechanisms. Near real-time AI-based propagation modelling provides the computational foundation to address them at the speed that autonomous systems require.

Read the full technical analysis

Forsk has published a white paper detailing the complete technical basis for this capability: the CNN architecture and training methodology, the GPU acceleration approach developed with NVIDIA, the full benchmark results across hardware configurations, and the accuracy validation against measurement data and calibrated models.