Abstract
The Allen-Cahn-Ohta-Kawasaki (ACOK) equation is a complex phase-field model characterized by nonlocal constraints and volume conservation requirements. Standard Physics-Informed Neural Networks (PINNs) often struggle with such global constraints. This paper introduces a modified PINN architecture that utilizes auxiliary neural networks and specialized loss functions to handle inverse Laplacians and volume constraints. By employing a multi-network strategy and temporal rescaling of collocation points, the proposed method achieves high-fidelity simulations of phase separation processes in 1D domains, overcoming the limitations of traditional PINN frameworks.
1. Introduction
Phase-field models are essential tools in materials science for simulating phase separation in substances like polymers or alloys. The Allen-Cahn-Ohta-Kawasaki (ACOK) equation represents a significant challenge due to its nonlocal nature, volume constraints, and zero-mean conditions. Standard PINNs, while effective for local differential equations, often fail to converge or violate physical laws when applied to the ACOK equation. This research proposes a modified PINN framework specifically engineered to address these "boss level" scientific machine learning problems.
2. Problem Formulation
The ACOK equation presents three primary difficulties for standard neural architectures:
- Nonlocality: The presence of an "inverse Laplacian" term creates global dependencies where the state at one point depends on the entire domain.
- Volume Constraints: The total amount of a specific material must remain constant over time.
- Zero-Mean Condition: A mathematical requirement that the average value of the inverse Laplacian term must be zero.
3. Methodology
The authors developed a multi-network strategy to reformulate the global integral problem into a local differential one that PINNs can handle via backpropagation.
Key components of the modified architecture include:
- Auxiliary Variable ($\nu$): A second output, $\nu$, is added to the main network ($Net_u$) to represent the inverse Laplacian. The relationship $-\Delta \nu = f(u) - \omega$ is enforced directly within the loss function.
- Net_v Sub-Network: A separate, shallow network handles volume constraints by taking time ($t$) as input and using a built-in uniform spatial mesh to calculate integrals accurately, overcoming sparsity issues common in random sampling.
- Weighted Loss Functions: Adjustable weights allow the optimizer to prioritize fundamental constraints, such as initial conditions, before fine-tuning complex dynamics.
- Two-Stage Optimization: The process utilizes the ADAM optimizer for an initial global search followed by L-BFGS-B for high-precision convergence.
Training data is generated using Latin Hypercube Sampling (LHS). To address accuracy degradation over time (the "tail problem"), the authors implemented polynomial temporal rescaling ($x^2$ or $x^3$), placing more training points at the start of the simulation where phase separation is most violent.
4. Results and Discussion
The modified PINNs achieved remarkably low error rates in 1D simulations for short time intervals (up to $1 \times 10^{-2}s$). The model successfully maintained the double-well structure and respected volume constraints that typically baffle standard PINNs. A key insight from the study is that architectural changes—rather than simple increases in network depth—are necessary to mirror the mathematical structure of nonlocal PDEs. Furthermore, temporal rescaling significantly improved the fit of the inverse Laplacian term, which is usually the first component to degrade in standard simulations.
5. Conclusion and Future Work
While the results in 1D are strong, the approach faces challenges in scaling to 2D due to increased computational overhead and hyperparameter sensitivity. Future research will explore "curriculum regularization"—training the model on easier, shorter intervals first—and the integration of advanced architectures like Fourier Neural Operators (FNOs) to tackle higher-dimensional problems more efficiently.
Summary updated