When a proton beam enters a patient's body, it slows down and generates a stream of neutrons, like a pebble thrown into water scattering droplets. Doctors use these neutrons to check at what depth the protons stopped, ensuring the irradiation hits the tumor precisely and spares healthy tissue.
But calculating such a "neutron rain" on a computer is a long story: a full simulation of a single beam can take hours. The researchers proposed replacing the slow calculation with a neural network that learns from existing examples and produces a neutron map in 23 seconds. The prediction error is about 7% for protons and 14% for neutrons, and the agreement of spatial distributions is nearly 100%.
Such a fast calculation will help not only verify irradiation accuracy but also estimate the dose from neutrons — important for patient safety. In the future, the method could be extended to other particles produced during irradiation.
Neutrons generated during proton therapy also irradiate the body themselves, so their distribution must be accounted for to avoid harming healthy tissues.