In a tank of pure water, particles traveling faster than light in that medium leave bluish flashes, like a plane breaking the sound barrier creates a sonic boom. From these flashes, you can tell what kind of visitor passed through: a gamma quantum or a neutron. This is no idle task: at borders and warehouses, you need to quickly distinguish a harmless radioactive cargo from weapon-usable material.

First, a simple physical threshold works. Gamma rays from cobalt-60 and cesium-137 give energies below 2.62 MeV, while neutrons from an americium-beryllium source give higher energies. Anything below the threshold is a pure gamma source, and no need to complicate things. But what if the energy is above the threshold, yet the source is still not neutron? That's where the second stage comes in.

Three algorithms (Bagging, CatBoost, and multilayer perceptron) look at the pulse shape over time. A neutron produces a more 'smeared' flash, a gamma a sharp one. The models vote, and the combined answer is accurate in 81.6% of cases, with an area under the ROC curve of 0.921. Such a team is more reliable than a single method.

Water is cheap and safe, and the algorithm is explainable: first physics, then machine learning. This matters for inspectors who must trust the device. The hybrid approach can be scaled to larger tanks and used in ports, airports, and nuclear facilities.