Precise radio coverage prediction is fundamental to cellular network planning. Before deploying a base station, operators must determine where signals will be strong and where buildings will create obstructions. Traditional approaches fall into two categories, each with inherent limitations. Empirical models, exemplified by the Hata model, express path loss as a closed-form function of distance and a few environmental parameters. While computationally efficient, these models neglect specific building geometries, producing identical predictions for equidistant transmitters regardless of whether one is obstructed.

Deterministic solvers, primarily ray tracing, simulate electromagnetic propagation through three-dimensional scenes to achieve high accuracy. However, their computational cost—minutes per map—is prohibitive for interactive applications or city-scale analysis.

This work introduces a third approach: training a neural network on simulator-generated maps to reproduce high accuracy in milliseconds while preserving geometric information. The resulting model achieves a root-mean-square error of approximately 1.7 dB on previously unseen cities—approximately 71% lower than a classical distance-based baseline—while operating at interactive speeds.


v1	0.052	~4.2
v2	0.034	~2.7
v3	0.030	~2.4
v4	0.021	~1.7
—	0.179	14.30