🔬 condensed matter

Training overdamped dynamics

This paper presents a framework for training overdamped many-particle systems to exhibit targeted rate-dependent mechanical behaviors by deriving local update rules based on the Rayleighian formulation and demonstrating their ability to tune properties like the viscous Poisson's ratio in disordered Maxwell materials.

Original authors: Marc Berneman, Daniel Hexner

Published 2026-02-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Marc Berneman, Daniel Hexner

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a giant, messy bowl of spaghetti. But instead of noodles, it's made of tiny springs and little shock absorbers (like the ones in a car) all tangled together. This is what scientists call a "soft material."

In the real world, things like honey, slime, your cells, or even the plastic in a water bottle behave like this. They don't bounce back instantly like a rubber ball (which is "inertia"). Instead, they move slowly and sluggishly, like trying to run through deep water. This is called overdamped dynamics.

The big question the authors asked was: Can we "teach" this messy spaghetti bowl to move in a specific, useful way?

Usually, if you want a material to stretch or squish in a certain way, you have to build it perfectly from scratch. But this paper proposes a different idea: Let the material learn on its own.

Here is how they did it, broken down into simple concepts:

1. The "Training" Concept: Directed Aging

Imagine you have a team of workers (the tiny shock absorbers in our spaghetti bowl) holding a heavy box.

The Old Way: You tell them exactly how to stand before they start.
The New Way (This Paper): You just tell them, "If you feel a lot of tension, get a little weaker."

The researchers created a rule where the shock absorbers get "tired" or "softer" the more they are stretched. Over time, the parts of the material that are working the hardest become the most flexible. This is called Directed Aging. It's like how your muscles get stronger with exercise, but in reverse: the parts under stress get "softer" to let the material flow in a specific direction.

2. The Goal: Teaching the Material to "Think"

The team wanted to train this material to do two tricky things:

The "Anti-Squish" (Negative Poisson's Ratio):
Normally, if you squeeze a sponge, it gets wider on the sides. If you pull it, it gets thinner.
The researchers trained their material to do the opposite. When they squeezed it, it got thinner on the sides. When they pulled it, it got wider. It's like a weird, magical sponge that defies normal physics. They did this by "aging" the shock absorbers until the material learned this strange behavior.
The "Remote Control" (Local Responses):
Imagine you have a giant trampoline with hundreds of people on it. You want to push down on one person in the corner, and have a person on the opposite side jump up.
Usually, the push just ripples out and fades away. But the researchers trained the material so that pushing on specific "source" points would cause a precise movement at a distant "target" point. It's like teaching the material to have a secret handshake where a nudge here causes a wave there.

3. The Secret Sauce: Two Speeds, Two Personalities

The coolest part of this discovery is that the material can have two different personalities depending on how fast you move it.

Fast Movement (The Elastic Personality): If you poke it quickly, it acts like a stiff spring. It remembers its shape.
Slow Movement (The Viscous Personality): If you push it slowly, it acts like thick honey. It flows and changes shape.

Because the "springs" and the "shock absorbers" are different parts of the system, the researchers could train them independently.

They could teach the springs to make the material stiff in one way.
They could teach the shock absorbers to make it flow in a completely different way.

This means they could create a material that is stiff when you move fast (like a solid) but soft and moldable when you move slow (like putty).

4. The "Equilibrium Propagation" Trick

They also used a second method called "Equilibrium Propagation." Think of this as a feedback loop.
Imagine you are trying to teach a dog to sit.

You ask it to sit (Free state).
It doesn't quite do it right.
You gently nudge it into the sitting position (Clamped state).
The dog "learns" the difference between what it did and what you wanted, and adjusts its behavior for next time.

The computer did this with the material, nudging it toward the desired shape and adjusting the "tiredness" of the shock absorbers until the material could do the trick perfectly on its own.

Why Does This Matter?

This is a huge step forward for smart materials.

In Medicine: Imagine a bandage that is stiff to protect a wound but soft and flexible when you move your arm slowly.
In Robotics: Soft robots that can be rigid when they need to lift something heavy but fluid when they need to squeeze through a small hole.
In Nature: It helps us understand how cells and tissues adapt to stress, essentially "learning" how to survive in their environment.

In a nutshell: The authors figured out how to "train" messy, slow-moving materials to remember specific shapes and behaviors, just like training a dog, but by letting the material's internal parts get "tired" and adapt over time. They turned a chaotic soup of springs and shock absorbers into a programmable, shape-shifting material.

1. Problem Statement

The paper addresses the challenge of controlling the dynamical responses of many-particle systems where inertia is negligible, a regime known as overdamped dynamics. This regime is ubiquitous in soft matter physics (polymers, colloids), biological systems (cellular mechanics), and active matter.

While static or quasi-static properties of materials can be tuned via inverse design or self-organization (e.g., directed aging), controlling time-dependent, rate-dependent behaviors in overdamped systems remains difficult. Existing methods often focus on quasi-static limits or Newtonian (inertial) dynamics. The authors aim to develop a framework to "train" overdamped systems to exhibit specific, programmable dynamical responses (such as specific viscous flows or rate-dependent elastic moduli) by autonomously adapting microscopic parameters without direct fabrication of the microstructure.

2. Methodology

The authors propose a theoretical framework based on the Rayleighian formulation of overdamped dynamics and apply it to a specific model system using two distinct training rules.

A. Theoretical Framework: The Rayleighian

Overdamped dynamics are governed by the equation $\sum_j \eta_{ij} \dot{x}_j = f_i$ , where $\eta$ is the dissipation matrix and $f$ are forces. The authors utilize the Rayleighian ( $R$ ), a function defined as:
$R = \frac{1}{2} \sum_{i,j} \eta_{ij} \dot{x}_i \dot{x}_j - \sum_i f_i \dot{x}_i$
The equations of motion are obtained by minimizing $R$ with respect to velocities ( $\partial R / \partial \dot{x}_i = 0$ ).

Key Insight: While the system trajectory does not extremize the time-integrated Rayleighian (unlike the action in Lagrangian mechanics), the principle of instantaneous extremization provides a foundation for constructing approximate learning rules. The Rayleighian is structurally analogous to elastic potential energy, where the dissipation matrix $\eta$ corresponds to the Hessian (stiffness), and velocities correspond to displacements.

B. Model System

The authors test their framework on a disordered 2D network of Maxwell viscoelastic elements.

Structure: A bonded network where each bond consists of a spring (stiffness $k_i$ ) and a dashpot (dissipation coefficient $\gamma_i$ ) in series.
Behavior: The system behaves elastically at short times (fast rates) and flows like a fluid at long times (slow rates).
Training Mechanism: The system is subjected to cyclic strain. The microscopic parameters ( $\gamma_i$ and potentially $k_i$ ) are updated based on local rules to minimize a cost function.

C. Training Rules

Two autonomous update rules are derived:

Directed Aging (Self-Organization):
- Inspired by physical aging and damage accumulation.
- Rule: The dissipation coefficient $\gamma_i$ degrades over time based on the local "creep" (rate of plastic strain).
- Equation: $\frac{d\gamma_i}{d\tau} = -\gamma_i r_{Age} (\frac{d\ell_{i,0}}{d\tau})^2$ .
- Effect: This creates a "soft direction" in the dissipation space, steering the system's flow response.
Equilibrium Propagation (Supervised Learning):
- Adapted from machine learning concepts for physical systems.
- Concept: Uses a "clamped" state (where the system is nudged toward a desired target behavior via an external cost function) and a "free" state.
- Rule: The gradient of a cost function $C$ with respect to $\gamma_i$ is estimated by the difference in the squared plastic strain rates between the clamped and free states:
  $\frac{\partial C}{\partial \gamma_i} \approx \frac{1}{2\beta} \left[ \left(\frac{d\ell^c_{i,0}}{d\tau}\right)^2 - \left(\frac{d\ell^f_{i,0}}{d\tau}\right)^2 \right]$
- Update: Parameters are updated via gradient descent using this estimated gradient.

3. Key Contributions

Framework for Overdamped Training: The first systematic derivation of learning/adaptation rules specifically for overdamped dynamics using the Rayleighian, bridging the gap between physical learning and dissipative systems.
Decoupling of Elastic and Viscous Control: Demonstrated that because elastic responses depend on spring constants ( $k_i$ ) and viscous responses on damping coefficients ( $\gamma_i$ ), these two regimes can be tuned independently. This allows for the creation of materials with distinct behaviors at fast vs. slow actuation rates.
Local vs. Global Control: Showed that training can be applied to global properties (like Poisson's ratio) and local, spatially specific responses (e.g., inducing strain at a distant target node via local actuation).

4. Results

The authors performed numerical simulations on disordered networks ( $N \approx 256-512$ nodes) to validate the rules:

Control of Viscous Poisson's Ratio:
- Using Directed Aging, they trained the system to exhibit a negative Poisson's ratio (auxetic behavior) at slow strain rates, while maintaining a positive ratio at fast rates.
- The crossover between elastic (positive $\nu$ ) and viscous (negative $\nu$ ) regimes is sharp and tunable.
- The distribution of inverse damping coefficients ( $1/\gamma_i$ ) evolved into a power-law distribution ( $P(1/\gamma) \sim (1/\gamma)^{-3}$ ), suggesting a self-organized critical state.
Local Response Training:
- The system was trained to produce a specific strain at a target node when multiple source nodes were actuated.
- Directed Aging: Successfully induced a target strain at slow rates, though the coupling decays with distance (Green's function limitation).
- Equilibrium Propagation: Achieved high-precision training of local responses. The error between the desired and actual target strain decreased by over an order of magnitude after training, demonstrating the efficacy of the supervised learning approach.
Multi-Functional Training:
- By simultaneously updating $k_i$ (for fast response) and $\gamma_i$ (for slow response), the authors trained a material to have a negative Poisson's ratio at fast rates and a positive Poisson's ratio at slow rates. This demonstrates full control over rate-dependent functionality.

5. Significance and Implications

Programmable Materials: The work suggests a pathway to create "smart" soft materials that can change their mechanical properties (stiffness, flow, auxeticity) based on the rate of loading, without requiring external power sources or complex internal machinery.
Biological Relevance: Since cellular and tissue mechanics are often overdamped, these principles could explain or engineer biological adaptation, such as how tissues remodel their stiffness or flow in response to mechanical stress.
Active Matter: The framework is applicable to active systems where local evolution rules are naturally coupled to dynamics.
Theoretical Advancement: It extends the concept of "physical learning" beyond quasi-static or inertial systems, providing a rigorous mathematical basis (via the Rayleighian) for how dissipative systems can "learn" to perform tasks.

In conclusion, Berneman and Hexner demonstrate that overdamped systems can be trained to exhibit targeted, rate-dependent behaviors through local, physically motivated update rules, opening new avenues for the design of adaptive soft and biological materials.