Coarse-Grained Dynamics with Spatial Disorder and Non-Markovian Memory

This paper introduces the spatial disorder-generalized Langevin equation (SD-GLE), a data-driven variational Bayesian method that successfully separates static spatial disorder from viscoelastic friction to accurately model anomalous diffusion and long-time dynamics in heterogeneous systems where standard approaches fail.

The Gist

Imagine you are trying to predict how a single drop of ink will spread through a cup of coffee.

If the coffee is perfectly smooth and uniform, the drop spreads out evenly and predictably. This is like the "standard" way scientists have tried to model how particles move in complex systems for a long time. They assume the environment is a smooth, average landscape.

But real life isn't like smooth coffee. Think of a hiking trail in a dense, foggy forest.

• The Forest (The System): It's full of hidden roots, rocks, and mud pits (spatial disorder).
• The Hiker (The Particle): They are trying to walk through it.
• The Mud (Viscoelastic Friction): The ground is sticky; the hiker sinks in and has to pull their leg out slowly.

The Problem with the Old Way

Previous methods tried to describe the hiker's journey by saying, "On average, the ground is a little sticky, and the hiker moves at a medium speed."

They ignored the fact that the hiker might get stuck in a specific mud pit for an hour, or get tripped by a specific root. Because they only looked at the "average," they missed the local traps.

When scientists tried to predict where the hiker would be tomorrow based on this "average" model, it failed. The model thought the hiker would eventually break free and run smoothly. In reality, the hiker is still stuck in that mud pit because the forest is full of unique, rugged obstacles that the "average" model smoothed over.

The New Solution: SD-GLE

The authors of this paper, Chuyi Liu and his team, invented a new tool called SD-GLE (Spatial Disorder-Generalized Langevin Equation).

Think of SD-GLE as a super-smart detective who doesn't just look at the average weather; they map every single pothole and puddle on the trail.

Here is how it works, using our hiking analogy:

1. Separating the "Trap" from the "Sticky":
   The old method confused the sticky mud (which slows you down temporarily) with the deep holes (which trap you for a long time). SD-GLE is smart enough to say: "Okay, this part of the delay is because the ground is sticky (memory), but this part is because the hiker fell into a specific hole (spatial disorder)." It untangles these two causes.

2. The "Random Map" (Gaussian Random Field):
   Instead of assuming the forest is flat, SD-GLE assumes the forest is a randomly generated 3D map. It uses a mathematical trick (Bayesian inference) to learn what that map looks like just by watching the hiker walk for a short time. It asks: "If the hiker got stuck here, there must be a hole here. If they moved fast there, the path must be clear."

3. Predicting the Long Hike:
   Because SD-GLE knows exactly where the "holes" and "rocks" are, it can predict the hiker's journey for days or weeks into the future.

   • The Old Model would say: "The hiker will eventually walk in a straight line." (Wrong!)
   • The SD-GLE Model says: "The hiker will get stuck in this specific cluster of rocks for a while, then bounce around in a weird pattern, and never quite reach the straight line." (Correct!)

Why Does This Matter?

This isn't just about hikers. This applies to:

• Living Cells: Proteins moving inside a cell are constantly bumping into a messy, crowded environment.
• Polymers and Glass: Materials that are hard to predict because their internal structure is chaotic.

By using this new "detective" method, scientists can finally predict how these complex systems behave over the long term using only short-term observations. It stops us from making the mistake of thinking a chaotic, messy system will eventually become smooth and predictable.

In a nutshell:
The old way tried to describe a messy forest by averaging out the trees and rocks. The new way (SD-GLE) draws a detailed map of every tree and rock, allowing us to accurately predict how someone will get lost (or find their way) in that forest for a very long time.

Technical Summary

1. Problem Statement

Predicting the long-term dynamics of particles in complex, heterogeneous systems (e.g., living cells, polymer networks, glass-forming liquids) is a central challenge in non-equilibrium statistical mechanics. These systems exhibit anomalous transport arising from the coupling of two distinct mechanisms:

1. Viscoelastic friction: A non-Markovian memory effect due to the environment's response.
2. Static spatial disorder: A rugged potential energy surface causing localized trapping and weak ergodicity breaking.

Limitations of Existing Methods:

• Traditional Phenomenological Models: Approaches like Fractional Brownian Motion (FBM) or Continuous-Time Random Walks (CTRW) typically assume a single underlying mechanism, failing to quantitatively reproduce coupled dynamics.
• Standard Generalized Langevin Equations (GLE): While data-driven GLEs can infer effective potentials and memory kernels, they rely on a global mean-field approximation. In heterogeneous environments, this approach averages out local spatial disorder. Consequently, the unresolved spatial heterogeneity is erroneously absorbed into the inferred memory kernel, leading to:
  • Overestimation of temporal correlations.
  • Failure to capture long-time trapping.
  • Unphysical extrapolation of long-time transport properties (e.g., predicting normal diffusion instead of sustained anomalous diffusion).

2. Methodology: The SD-GLE Framework

The authors introduce the Spatial Disorder-Generalized Langevin Equation (SD-GLE), a data-driven framework designed to explicitly disentangle static spatial disorder from viscoelastic memory using a variational Bayesian inference approach.

Key Technical Components:

• Gaussian Random Field Prior: The local effective potential $U(Q)$ is modeled not as a fixed function but as a realization of a Gaussian Random Field (GRF), $U(Q) \sim \mathcal{GP}(U_{\text{mean}}, C(Q, Q'))$. This captures the statistical nature of spatial disorder.
• Variational Inference with Inducing Points: To handle the high-dimensional path integral of the potential, the authors employ a sparse Gaussian Process approximation. They introduce a set of learnable "inducing variables" ($u$) to create a tractable finite-dimensional representation of the potential landscape.
• Marginalization Strategy:
  • Spatial Marginalization: The likelihood is maximized by integrating over all possible potential configurations using the Evidence Lower Bound (ELBO). This allows the model to adaptively capture local spatial fluctuations via single-trajectory variational posteriors.
  • Temporal Marginalization: The non-Markovian friction is embedded into an extended phase space using auxiliary variables ($\zeta$) representing heat bath degrees of freedom. This transforms the problem into a Markovian process, allowing the use of a Kalman filter (predictor-corrector) to compute the exact trajectory likelihood efficiently.
• Optimization: The model parameters (spatial correlation function $\theta_C$, memory kernel parameters $\theta_K$, and posterior distributions) are optimized by maximizing the ELBO, which combines the trajectory likelihood and the Kullback-Leibler (KL) divergence between the approximate and prior distributions.

3. Key Contributions

1. Explicit Decoupling: The framework successfully separates static spatial correlations from dynamic viscoelastic memory, a task previous GLE methods failed to perform quantitatively.
2. Recovery of Rugged Landscapes: Unlike mean-field approaches that yield featureless potentials, SD-GLE reconstructs the rugged effective potential energy surface, explicitly resolving local trapping environments.
3. Accurate Long-Time Extrapolation: By preserving the static disorder, the model correctly extrapolates long-time dynamics from short trajectories, avoiding the artificial transition to normal diffusion seen in standard GLEs.
4. Capturing Dynamic Heterogeneity: The framework recovers essential features of disordered transport, including transient non-Gaussian displacement distributions and massive inter-trajectory dispersion caused by weak ergodicity breaking.

4. Results and Validation

The authors validated SD-GLE on two systems:

A. Ideal 1D System:

• Setup: Trajectories generated from a known Gaussian random field potential with varying disorder strengths ($\sigma$).
• Findings:
  • Memory Kernel: Standard GLE overestimated the memory kernel amplitude by absorbing spatial disorder. SD-GLE accurately recovered the true non-Markovian memory.
  • Potential Reconstruction: SD-GLE faithfully tracked the reference rugged landscape, whereas standard GLE produced a smooth, unphysical potential.
  • MSD Prediction: Standard GLE failed to sustain anomalous diffusion, transitioning prematurely to Fickian diffusion. SD-GLE maintained the correct anomalous diffusion behavior over time scales exceeding $10^4$, with significantly lower relative prediction errors across all disorder strengths.

B. 2D Kob-Andersen (KA) Glass-Forming Liquid:

• Setup: Molecular Dynamics (MD) simulations of a binary mixture with pinned particles to induce disorder.
• Findings:
  • Non-Gaussianity: In supercooled states, particle displacements exhibited heavy tails (non-Gaussian). SD-GLE quantitatively reproduced the non-Gaussian parameter $\alpha_2(t)$, while standard GLE (governed by stationary Gaussian noise) failed ($\alpha_2 \approx 0$).
  • Ergodicity Breaking: The model successfully reproduced the time-averaged Mean Square Displacement (MSD) dispersion and the Ergodicity Breaking (EB) parameter observed in MD data. Standard GLE predicted purely ergodic behavior, failing to capture the weak ergodicity breaking inherent to the disordered system.

5. Significance

The SD-GLE framework represents a paradigm shift in coarse-grained modeling for heterogeneous systems.

• Physical Insight: It demonstrates that static spatial disorder is not merely a geometric correction but a fundamental physical requisite for recovering dynamic heterogeneity and anomalous transport.
• Predictive Power: It enables the construction of predictive, data-driven dynamical models that can extrapolate long-term behavior from limited, short-time trajectory data.
• Scalability: The use of variational inference and inducing points ensures the method is computationally tractable and scalable, providing a robust foundation for modeling complex systems like biological cells and glassy materials where traditional mean-field approximations fail.