Fisher’s information matrix approach for Fourier features physics-informed neural networks for two-dimensional local time-fractional anomalous diffusion equations with nonlinear thermal diffusivity

Purpose

This study aims to explore the influence of Fourier-feature enhanced physics-informed neural networks (PINNs) on effectively solving two-dimensional local time-fractional anomalous diffusion equations with nonlinear thermal diffusivity. By tackling the shortcomings of conventional numerical methods in managing fractional derivatives and nonlinearities, this research addresses a significant gap in the literature regarding efficient solution strategies for complex diffusion processes.

Design/methodology/approach

This study uses a quantitative methodology featuring a feed-forward neural network architecture combined with a Fourier feature layer. Automatic differentiation is implemented to ensure precise gradient calculations for fractional derivatives. The effectiveness of the proposed approach is showcased through numerical simulations across various sub-diffusion and super-diffusion scenarios, with fractal space parameters adjusted to examine behavior. In addition, the training process is assessed using the Fisher information matrix to analyze the loss landscape.

Findings

The results demonstrate that the Fourier-feature enhanced PINNs effectively capture the dynamics of the anomalous diffusion equation, achieving greater solution accuracy than traditional methods. The analysis using the Fisher information matrix underscores the importance of hyperparameter tuning in optimizing network performance. These findings support the hypothesis that Fourier features improve the model’s capacity to represent complex solution behaviors, providing the relationship between model architecture and diffusion dynamics.

Originality/value

This research presents a novel approach to solving fractional anomalous diffusion equations through Fourier-feature enhanced PINNs. The results contribute to the advancement of computational methods in areas such as thermal engineering, materials science and biological diffusion modeling, while also providing a foundation for future investigations into training dynamics within neural networks.