This paper presents an approach based on the gamma-ray transmission technique
and artificial neural network for accurately measuring the thickness of
various materials in flat sheet form. The gamma-ray transmission system
comprises a NaI(Tl) scintillation detector coupled with a 137Cs radioactive
source. The artificial neural network model predicts the sample thickness
through three input features: mass density, linear attenuation coefficient,
and ln(R) - where R represents the ratio of areas under the 662 keV peak in
spectra acquired from measurements with and without the sample. The
artificial neural network model was trained using simulation data generated
by MCNP6 code, facilitating the creation of comprehensive datasets covering
diverse material types and thickness variations at a low cost.
Hyperparameters of the artificial neural network model were defined by
several optimization methods, such as hyperband-bayesian, tree-structured
Parzen estimator, and random search, to establish an optimal artificial
neural network architecture. Subsequently, the optimal artificial neural
network model was deployed to predict the thickness of graphite, aluminum,
copper, steel, and polymethyl methacrylate sheets, using input data obtained
from the experiments. The results showed a good agreement between predicted
and reference thicknesses, with a maximum relative deviation of 1.94 % and
an average relative deviation of 0.52%.