Divide and conquer!

Not all people can do all jobs, and not all models work well in all scenarios.

I was recently doing some looking into a shallow water problem with dBSea noise modelling software, to see how well it deals with the scenario. You might wonder why shallow water is a problem, surely that's a simple environment!? 

No, shallow water problems are notorious, as noise propagation is governed by a combination of acoustic ducting between surface and bottom and complex and numerous reflections of the sea floor and sea surface.

No single model does this well, so I mixed the relatively simple model of spherical and cylindrical spreading with ray-tracing for something that turned out to be very close to measured, real life values (±4 dB at 5 km). DOSITS has a great article on geometrical spreading. 

My "Spherical+Cylindrical" is simply assuming spherical spreading loss from the source to the range that equals the depth. From then on I assume that the sound spreads cylindrically (because the surface and bottom reflects sound). I then apply some frequency dependent attenuation (see why in "Wave Length Matters"), the folks from National Physics Laboratory have a nifty calculator for this (easier than figuring out the physics for yourself).

Figure 1. Measurements minus predicted values. In other words, how many dB is the real world louder than the prediction.

The ray tracing is pretty much what it says on the label; you trace a single ray at a time on it's journey through the ocean, calculating refraction and dampening stepwise as you go along (read more here). Then you do another one, and another one, and another one, and ano... you get the picture.

So by splitting up the scenario into smaller portions, each algorithm can crunch the numbers it's best suited to - and voilà, you've conquered the noise.

For the observant geek:

How to know what models to use for what frequencies if you don't have any real world data to check it against?

hmmm... good one - turns out the sort of plot I used here (Figure 1) shows abrupt changes in polarity (values go from negative to positive quickly), if you've chosen the wrong crossover-frequency. This is only a guideline, but works surprisingly well is the source spectrum is smooth looking.

An example of how to choose your crossover frequency between solvers better.

Bear in mind that there's no mathematical reasoning on my part behind this - I have just found it to work in practise - use at your own risk!

Thanks for reading this post - please comment if you like.