The paper describes a novel neural-network-based approach to study the distributions of secondaries produced in hadronic showers using observables provided by highly granular calorimeters. The response is analysed of the highly granular scintillator-steel hadron calorimeter to negative pions with momenta from 10 to 80 GeV simulated with two physics lists from the Geant4 package version 10.3. Several global observables, which characterise different aspects of hadronic shower development, are used as inputs for a deep neural network. The network regression model is trained using a supervised learning and exploiting true information from the simulations. The trained model is applied to predict a number of neutrons and energy of neutral pions produced within a hadronic shower. The achieved performance and possible application of the model to validation of simulations are discussed.