Overview
This study investigates a data-driven methodology for predicting coarse-grained dynamics within multiscale partial differential equation (PDE) systems. The approach adopts a closure-free operator-learning perspective, focusing on the direct learning of a surrogate evolution operator for resolved fields. This learning process is based on filtered high-fidelity trajectories.
Research Context
The work is motivated by the Mori-Zwanzig formalism, which provides a theoretical framework for understanding the evolution of coarse-grained variables. The objective is to develop a computational tool that can learn the complex, non-linear relationships governing coarse-grained dynamics without explicit closure modeling, which is often a challenge in multiscale simulations.
Approach
The proposed methodology involves a spatiotemporal neural operator designed to map a resolved history slab to a resolved future slab. Specifically, this operator takes a history slab defined over a spatial domain $\Omega$ and a time interval $[-T_{\mathrm{\in}}, 0]$ and maps it to a future slab over $\Omega$ and $[0, T_{\mathrm{out}}]$.
- Spatial Mixing: Fourier convolution is employed for handling spatial mixing within the operator.
- Temporal Mixing: A causal kernel operator is used for temporal mixing. This operator incorporates position-attention weights on time lags, which are designed to encode finite-memory effects characteristic of resolved dynamics. This causal structure also preserves the directionality inherent in a history-to-future mapping.
- Inductive Bias: To enhance the robustness of autoregressive rollouts and mitigate nonconservative artifacts, the method embeds a flux-form inductive bias. This is achieved by parameterizing the windowed update mechanism in an explicit divergence form.
- Memory Length Selection: A data-driven guideline is provided for determining the optimal memory length $T_{\mathrm{\in}}$. This guideline relies on computing the decorrelation time of a closure-injection diagnostic derived from filtered trajectories.
Findings
The proposed flux-form spatiotemporal neural operator was validated across several test cases involving coarse-grained dynamics:
- Viscous Burgers' equation
- Kuramoto-Sivashinsky equation
- Two-dimensional turbulent flows
Across these validation cases, the operator demonstrated stable autoregressive rollouts. It also yielded improvements in long-horizon accuracy and statistical fidelity compared to conventional approaches or models lacking the embedded flux-form inductive bias.
Why This Matters
The development of a closure-free, data-driven operator for coarse-grained dynamics addresses a fundamental challenge in simulating multiscale PDE systems. By directly learning the evolution operator from high-fidelity trajectories and incorporating inductive biases like the flux-form, the method offers a pathway to more stable and accurate long-term predictions in complex physical systems without relying on potentially incomplete or inaccurate closure models.
Potential Applications
While specific applications are not detailed, the capability to accurately predict coarse-grained dynamics in complex PDE systems suggests relevance for simulations in fluid dynamics, turbulence modeling, and other areas where multiscale phenomena are prevalent and traditional modeling approaches face challenges related to closure and long-term stability.