TL;DR
- Empirical models are fast and need calibration; ray-tracing models are slower and need good geometry.
- Calibration against CW measurements in each clutter type matters more than the choice between model families.
- Never compare model outputs to drive test RSRP without accounting for the difference between CW and live-network measurements.
What is the difference between an empirical model and ray-tracing?
An empirical model predicts path loss from distance, frequency, antenna height and a clutter correction derived from measurements elsewhere. It is fast, needs little input data, and its accuracy depends almost entirely on how well its clutter corrections match your environment. Aster is in this family.
A ray-tracing model computes actual propagation paths against building geometry, including reflections and diffraction. CrossWave is in this family. It handles street canyons and dense urban geometry far better than any empirical correction can, but it needs accurate 3D building data and it computes more slowly.
| Situation | Sensible default | Reason |
|---|---|---|
| Rural macro planning | Empirical, calibrated | Simple geometry, calibration sufficient |
| Suburban / small town | Empirical, calibrated per clutter | Mixed clutter, fast iteration matters |
| Dense urban | Ray-tracing | Geometry dominates; corrections cannot capture it |
| Street canyon / CBD | Ray-tracing with 3D data | Reflections and diffraction are the signal path |
| No 3D data available | Empirical + heavy calibration | Ray-tracing without geometry is worse than useless |
When does the model choice actually matter?
In rural and suburban areas with modest clutter, a calibrated empirical model is usually within a usable tolerance and costs a fraction of the effort. The geometry is simple enough that ray-tracing adds precision you cannot validate anyway.
In dense urban areas the ranking reverses. Street canyons, courtyard buildings and mixed heights produce propagation that no distance-based correction captures, and the difference between two candidate sites is often entirely a geometry effect. This is where ray-tracing earns its cost.
How do you calibrate a model properly?
Collect CW measurements along routes that cross each clutter type, then fit the model's correction parameters so predicted and measured path loss agree. The discipline that matters is separating the data: calibrate on one set of routes and validate on a different set. Fitting to the same data you validate against will always look good and tells you nothing.
Keep the calibration per clutter type rather than global. A single correction that splits the difference between dense urban and open rural will be wrong in both, and the error will be largest exactly where the design decisions are hardest.
Why do model predictions disagree with drive test results?
Because they are measuring different things. CW measurements are typically taken with an omnidirectional antenna at a known height along a controlled route, while drive test RSRP comes from a live network with real antenna patterns, real tilt, real load and real handovers. A 6 to 10 dB gap between the two is normal and is not evidence that the model is wrong.
The correct comparison is model to CW, then CW to live network with the antenna and load differences accounted for. Comparing model output directly to drive test RSRP and then adjusting the model to close the gap produces a model calibrated to the wrong target.