Abstract
Fractional differential equations (FDEs) are essential for modeling systems with memory effects and non-local dynamics, yet they present significant computational challenges. Traditional Physics-Informed Neural Networks (PINNs) struggle with FDEs because standard automatic differentiation is designed for integer-order derivatives. This paper discusses the Physical Model-driven Neural Network (PMNN), a framework that integrates numerical discretization schemes—specifically L1 and L2-1σ approximations—directly into deep learning architectures. By bridging numerical analysis and neural networks, PMNN provides a robust solver for time-fractional equations with high accuracy and efficiency.
1. Introduction
In fields ranging from viscoelasticity to signal processing and anomalous diffusion, fractional differential equations (FDEs) have become the standard mathematical tool for describing "non-local" processes. Unlike standard differential equations, where the state of a system depends only on its immediate surroundings and current time, FDEs account for the historical trajectory of the system, capturing essential memory effects. However, the non-local nature of fractional operators makes analytical solutions rare and numerical solutions computationally demanding.
2. The Challenge of Fractional Calculus in Deep Learning
The emergence of Physics-Informed Neural Networks (PINNs) has transformed the landscape of partial differential equations (PDEs). By leveraging automatic differentiation (AD), PINNs can solve complex equations without traditional meshes. However, AD engines in popular libraries like PyTorch and TensorFlow are fundamentally built for integer-order calculus (1st, 2nd derivatives, etc.). They cannot natively compute a derivative of order 0.5 or 0.75. Previous attempts to adapt PINNs to fractional calculus often resulted in high computational overhead or a failure to strictly adhere to the underlying physical constraints.
3. Methodology: The PMNN Framework
The Physical Model-driven Neural Network (PMNN) addresses these limitations through a "discretize-then-learn" strategy. Instead of relying on AD for the fractional component, PMNN embeds established numerical interpolation techniques as a "bridge" within the neural network's loss function.
3.1 Semi-Discretization Schemes
The framework utilizes two primary discretization schemes to approximate the temporal fractional operator:
- L1 Approximation: A standard piecewise linear interpolation method with a theoretical convergence order of 2 - α.
- L2-1σ Approximation: A more sophisticated superconvergent scheme that achieves a higher convergence order of 3 - α.
3.2 Neural Iteration and Loss Function
The core innovation lies in substituting the neural network's output into these discretized schemes to form a temporal iteration. The loss function is then defined by the residual of this iteration scheme, combined with the initial and boundary conditions. By minimizing this loss, the network learns to satisfy the discretized physics of the equation across the entire spatio-temporal domain.
4. Experimental Validation
The PMNN framework was validated using three synthetic benchmarks where exact analytical solutions are known, allowing for rigorous error analysis:
- Example 1: A single-term temporal fractional ordinary differential equation (FODE).
- Example 2: A 1D time-fractional convection-diffusion equation (FPDE).
- Example 3: A 2D time-fractional convection-diffusion equation (FPDE).
5. Key Results and Discussion
The experiments yielded several critical insights into the performance and efficiency of PMNN:
- Superiority of L2-1σ: The L2-1σ variant consistently outperformed the L1 variant. Due to its higher-order convergence, it reached lower error floors—often an order of magnitude better—and converged significantly faster, particularly for higher fractional orders (α = 0.75).
- High Precision: In the 2D FPDE benchmark, the PMNN (L2-1σ) achieved an L2 relative error as low as 4.17 × 10⁻⁵.
- Spatial Efficiency: A notable finding was the model's robustness to sparse spatial data. Increasing the number of spatial training nodes beyond a minimal threshold (e.g., Nx = 11) did not significantly improve accuracy, suggesting that PMNN can be trained efficiently without dense spatial grids.
- Stability: The framework remained stable across the entire range of fractional orders α ∈ (0, 1), accurately preserving the sub-diffusion characteristics inherent in the physical models.
6. Conclusion and Future Directions
PMNN represents a significant step forward in combining the flexibility of deep learning with the rigor of numerical analysis. By using DNNs as a bridge for interpolation techniques, the framework overcomes the inherent limitations of automatic differentiation in fractional calculus.
While the current model is tailored for single-term temporal fractional equations, future work will focus on extending PMNN to handle multi-term temporal fractional equations and spatial fractional derivatives. Additionally, the authors aim to explore the framework's performance on noisy, real-world experimental data and investigate the theoretical basis for observed error symmetries in the model's output.
Summary updated