COMPARATIVE ANALYSIS OF NEURAL NETWORK METHODS FOR SOLVING THE MODEL OF A SINGLE DISLOCATION SOURCE OF GEOACOUSTIC EMISSION

This paper presents a comparative analysis of computational methods for modeling a single dislocation source of geoacoustic emission radiation, the Berlage oscillator. The paper considers the Rosenbrock numerical method of the fourth order of accuracy, a parallel modification of a physically informed neural network into three subdomains (TriDomainPINN), as well as a neural network based on Legendre polynomials (LeNN). The parallel version of PINN makes it possible to increase the efficiency of the solution and reduce the learning time by decomposing the modeling area. All algorithms are implemented in Python using the PyTorch library; the implementation provides a mechanism for early stopping of network training and the distribution of collocation points according to Chebyshev’s law. Accuracy, convergence rate, and computational efficiency were evaluated for each method. Numerical experiments show that the Legendre polynomial network provides high approximation accuracy with a significant reduction in computational costs and time.