Akademik Veri Yönetim Sistemi
Engineering Science and Technology, an International Journal, cilt.80, ss.1-23, 2026 (SCI-Expanded)
In neural networks, the topology is established based on experience and trial and error methods. In this study, a new method called Diffraction Analysis is proposed for systematically forming a neural network’s topology for the first time. With the proposed method, the network’s hidden layers and neuron structures can be constructed from scratch. Additionally, the network’s weight values can be determined using the proposed mathematical approach with a single iteration. Unlike conventional neural networks that rely on backpropagation, iterative training procedures, and hyperparameter optimization, the proposed approach determines the network topology from scratch and resolves all weight values analytically in a single pass, eliminating the training phase entirely. To facilitate reproducibility and transparency, the construction of the topology and the determination of weight values are demonstrated step-by-step through illustrative examples. The performance of Diffraction Analysis is tested with the Sinx2 dataset for the regression problem, with the Iris, Diabetes, Heart Disease and Mobile Price datasets for the classification problem. Experimental results demonstrate that the proposed method achieved accuracy rates of 95% on the Iris dataset, 74% on the Diabetes dataset, 99% on the Heart Disease dataset, and 83% on the Mobile Price dataset, respectively, with highly efficient computational times ranging from 1 to 81 ms. To comprehensively validate the robustness of the algorithm, standard error and classification metrics (MAE, MSE, RMSE, Accuracy, Precision, Specificity, Sensitivity) are presented alongside ROC curves, confusion matrices, and t-SNE plots. At the same time, the method demonstrates competitive performance on well-established benchmark datasets In this way, the model presents a solution approach to the black-box problem of artificial neural network studies. It also implies that the transformation of neurons in the output layer can convert regression problems into classification problems. This method presents an alternative and systematic solution approach to the topology construction problem in artificial neural networks.