Prediction of luminescence curves of quartz mineral with artificial neural networks

In this thesis study, the thermoluminescence (TL) and optically stimulated luminescence (OSL) properties of quartz mineral, separated from marine sand samples collected from a public beach in Erdemli, Mersin (Turkey), were comprehensively investigated, and the obtained data were modeled using artificial neural networks (ANN). In the experimental phase, optical filter selection, preheat temperature optimization, dose–response relationships, heating rate effects, and reusability tests were performed. TL measurements revealed three distinct regions: a linear region between 1–100 Gy, a sublinear region between 100–360 Gy, and a supralinear region above 360 Gy. For OSL measurements, a linear region was observed between 0.1–20 Gy, a sublinear region between 20–520 Gy, and a saturation behavior beyond 520 Gy. In the modeling phase, multilayer feed-forward ANN models were developed in MATLAB using the Neural Net Fitting (NNF) interface with three different optimization algorithms: Levenberg–Marquardt (LM), Bayesian Regularization (BR), and Scaled Conjugate Gradient (SCG). The LM algorithm provided the highest correlation coefficient (R = 0.999) and the lowest mean squared error (MSE) values in low- and medium-dose ranges, while the BR algorithm effectively prevented overfitting at higher doses, resulting in more balanced predictions. The SCG algorithm, although advantageous in terms of low memory consumption and shorter training time, exhibited a tendency toward overfitting at lower doses and showed limited generalization capability. The overall results demonstrate that the TL and OSL properties of quartz mineral hold strong potential for dosimetric applications and that ANN-based modeling can predict luminescence curves with high accuracy. This approach is expected to shorten experimental procedures, increase data production efficiency, and contribute significantly to OSL dating and radiation dosimetry studies.