🌀 nonlinear sciences

Hebbian Physics Networks: A Self-Organizing Computational Architecture Based on Local Physical Laws

The Hebbian Physics Network (HPN) introduces a self-organizing computational framework where physical states and transport operators co-evolve through local, residual-driven Hebbian adaptation, enabling the intrinsic emergence of thermodynamically stable and physically consistent transport geometries without global optimization or rigid grids.

Original authors: Gunjan Auti, Hirofumi Daiguji, Gouhei Tanaka

Published 2026-03-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Gunjan Auti, Hirofumi Daiguji, Gouhei Tanaka

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 are trying to organize a chaotic city during a massive power outage. The streets are clogged, people are panicking, and resources (like food and water) are stuck in the wrong places.

The Old Way (Classical Physics):
Traditionally, scientists act like a rigid city planner. They draw a fixed map with permanent roads and traffic lights. They say, "No matter what happens, traffic must flow exactly according to this pre-drawn map." If a road gets blocked, the planner tries to force the traffic through it anyway, or they calculate a new global route for the whole city at once. It's like trying to solve a giant puzzle by forcing every piece into a pre-cut slot. It works well for calm days, but when the city is in chaos (like a storm or a traffic jam), the rigid map doesn't fit the reality.

The New Way (Hebbian Physics Networks - HPN):
The paper introduces a revolutionary idea: What if the roads themselves could change shape?

Instead of a fixed map, imagine the city is made of smart, living roads. These roads are made of a special "plastic" material that can stretch, shrink, or disappear based on where the traffic is stuck.

Here is how the Hebbian Physics Network (HPN) works, using simple analogies:

1. The "Stuck Traffic" Signal (The Residual)

In this new system, every intersection has a sensor. If too much traffic piles up at an intersection (a "residual"), the sensor screams, "We are unbalanced!"

Old Way: You ignore the scream and keep driving on the broken road.
HPN Way: The scream is the most important thing. It tells the road exactly where to change.

2. The "Plastic Roads" (Adaptive Weights)

The connections between intersections (the edges of the network) are not fixed. They are weights.

If a road is helping to clear a traffic jam, the HPN makes that road wider and smoother (increasing the weight).
If a road is making the jam worse, the HPN narrows it or closes it (decreasing the weight).

This happens locally. The road at Intersection A doesn't need to know what's happening at Intersection Z. It only listens to the traffic jam right in front of it and adjusts itself instantly.

3. The "Anti-Hebbian" Rule (Learning by Unlearning)

The paper uses a fancy term called "Anti-Hebbian learning." In simple terms, think of it as "Learning by fixing mistakes."

In normal learning (like studying for a test), you repeat what you know to get stronger.
In this physics network, the system learns by removing the things that cause stress. If a connection causes a traffic jam, the system "unlearns" that connection. It prunes the bad paths until only the efficient paths remain.

4. The Magic Result: Self-Organization

The most amazing part is that no one draws the final map.

You start with a random mess of roads and a chaotic city.
You just let the "traffic jams" (residuals) drive the changes.
Over time, the roads naturally rearrange themselves into the perfect highway system to move the traffic.
The "laws of physics" (like how water flows or how heat spreads) emerge from this process. The system discovers the laws of physics by trying to stop the traffic jams.

A Real-World Analogy: The Grass Field

The authors use a great analogy in the paper:

Top-Down (Old Way): You see grass bending in the wind. You assume you know the laws of wind (Navier-Stokes equations) and try to calculate exactly how the wind must be blowing to bend the grass that way.
Bottom-Up (HPN Way): You don't know the wind laws. Instead, you look at each blade of grass. You ask, "How does this blade react to the grass next to it?" You let every blade adjust its position based on its neighbor. Eventually, the whole field of grass organizes itself into a perfect pattern that looks like it was blown by a specific wind, but the wind pattern was never calculated—it just emerged from the local adjustments.

Why Does This Matter?

It's Robust: If a road breaks, the system doesn't crash. It just reroutes the "plastic" roads around the break instantly.
It's Local: It doesn't need a supercomputer to calculate the whole city's traffic at once. Every intersection solves its own problem.
It Works in Chaos: It's perfect for turbulent flows, complex chemical reactions, or anything where the rules aren't clear beforehand. The system figures out the rules as it goes.

Summary

The Hebbian Physics Network is a computer program that stops trying to solve equations on a fixed grid. Instead, it builds a living, breathing network where the connections between points change shape to fix local problems. By constantly trying to "fix the jams" locally, the entire system naturally organizes itself into a perfect, efficient flow, discovering the laws of physics along the way.

It's like teaching a city to fix its own traffic jams by letting the roads grow and shrink until the traffic flows perfectly.

1. Problem Statement

Classical computational physics relies on a fundamental abstraction: the separation of state (e.g., velocity, temperature) and structure (the geometric transport operator). In traditional methods (Finite Difference, Finite Element, etc.), the constitutive operator governing transport is fixed a priori on a rigid grid. Conservation laws are enforced by applying this fixed operator globally, often requiring synchronized time-marching and global solvers (e.g., Poisson solvers for incompressibility).

While successful, this approach assumes the transport geometry is static, even in far-from-equilibrium regimes where physical systems naturally self-organize. Existing adaptive methods (mesh refinement, coefficient tuning) modify parameters within a fixed topology but do not allow the transport structure itself to evolve. The authors argue that from the perspective of non-equilibrium thermodynamics, the constitutive operator should be a dynamic field that co-evolves with the state to restore local conservation, rather than a static background object.

2. Methodology: The Hebbian Physics Network (HPN)

The authors propose the Hebbian Physics Network (HPN), a computational framework where the transport geometry is plastic and adapts locally to restore broken conservation laws.

Architecture: The system is defined on a directed graph $G=(V, E)$ $G = (V, E)$ .
- Nodes ( $i$ ): Store intensive state variables $U_i$ (e.g., concentration, momentum density) and have a local capacity $\Omega_i$ .
- Edges ( $j \to i$ ): Carry adaptive coupling weights $W_{ij}$ representing the conductance of the transport pathway.
- Signal ( $S_{ij}$ ): A state-centered signal (typically $S_{ij} = U_j$ ) transmitted from node $j$ to $i$ .
Flux and Residuals:
- Flux is defined as $J_{ij} = W_{ij}S_{ij}$ .
- The Residual ( $R_i$ ) is the local violation of conservation (e.g., divergence of flux) at node $i$ : $R_i = \sum (J_{ij} - J_{ji})$ . This acts as a local thermodynamic force.
Dynamics:
- State Evolution: The state $U_i$ relaxes to reduce the residual $R_i$ at a rate determined by local capacity: $\Delta U_i \propto R_i / \Omega_i$ .
- Structural Adaptation (The Core Innovation): The weights $W_{ij}$ $W_{ij}$ evolve via a residual-modulated anti-Hebbian rule:
  $\Delta W_{ij} = -\eta_w f(R_i)S_{ij} + \epsilon g(|R_i|)S_{ij} - \lambda W_{ij}$
  - Plastic Yield ( $-\eta_w f(R_i)S_{ij}$ ): Drives adaptation. If a pathway correlates with local accumulation (imbalance), its conductance is adjusted to relieve stress. This term constitutes a local gradient descent on the residual energy.
  - Promotion ( $\epsilon g(|R_i|)S_{ij}$ ): A small baseline permeability preventing pathways from vanishing completely near equilibrium.
  - Decay ( $-\lambda W_{ij}$ ): Prevents unbounded growth of weights.
Thermodynamic Consistency: The framework ensures that global conservation is an exact structural invariant ( $\sum R_i = 0$ ) at all times, while local imbalances drive the system toward a state where the transport operator satisfies Onsager's reciprocal relations (symmetry) near equilibrium.

3. Key Contributions

Unified Co-Evolution: HPN treats the constitutive geometry ( $W$ ) and the physical state ( $U$ ) as a coupled dynamical system. The "equation" is not solved; the system relaxes into a solution where the geometry itself adapts to satisfy conservation.
Local Gradient Descent: The authors prove that the anti-Hebbian weight update rule corresponds to a strictly local gradient descent on a residual energy functional ( $L = \frac{1}{2} \sum R_i^2 / \Omega_i$ ). This guarantees thermodynamic admissibility.
Emergence of Onsager Reciprocity: In the near-equilibrium limit, the adaptive weights spontaneously organize into a symmetric, positive-definite form, recovering Onsager's reciprocal relations without explicit enforcement.
No Global Solvers: The method eliminates the need for global matrix inversions (e.g., for pressure in incompressible flow) or global time-marching. Constraints are enforced locally through the evolution of the weights.
Self-Organizing Topology: The network learns the effective transport geometry (e.g., diffusion paths or flow streamlines) solely from random initial conditions driven by local residuals.

4. Results

The framework was demonstrated on two canonical problems:

Scalar Diffusion:
- Setup: An unstructured graph with random initial weights and a localized concentration perturbation.
- Outcome: The network evolved a transport geometry that reproduced parabolic diffusion fronts ( $r \propto \sqrt{t}$ ).
- Observations: Global conservation error remained at machine precision ( $O(10^{-11})$ ). The Lyapunov function (residual energy) decreased monotonically. The weight distribution became bistable (pruning inactive pathways), effectively learning the diffusion operator.
Incompressible Lid-Driven Cavity Flow:
- Setup: A hybrid formulation where the Navier-Stokes operator is used only to compute residuals (momentum imbalance and continuity violation), while HPN adapts the transport weights. Pressure acts as a Lagrange multiplier to enforce incompressibility locally.
- Outcome: Starting from a quiescent state, the network self-organized into the characteristic recirculating vortex structure of the cavity.
- Dynamics: The system exhibited a bifurcation where the transport operator reorganized to handle transverse momentum redistribution. The learned weights developed heavy-tailed distributions corresponding to shear layers.
- Validation: The converged steady-state velocity profiles at Reynolds number $Re=100$ matched standard benchmark data (Ghia et al.) with high accuracy.

5. Significance and Implications

Reframing Computation: HPN shifts the paradigm from "solving a fixed equation" to "thermodynamic relaxation." The computation is the relaxation process where state and geometry co-evolve.
Physics-Informed Learning: Unlike standard Physics-Informed Machine Learning (PIML) which imposes physics via global loss functions or fixed operators, HPN embeds physics bottom-up. The system learns the constitutive law by minimizing local thermodynamic stress.
Far-from-Equilibrium Applications: The framework is particularly relevant for complex, non-equilibrium systems (e.g., turbulence, active matter) where effective transport laws are unknown or state-dependent. It allows the discovery of effective transport topologies dynamically.
Geometric Interpretation: The authors draw a conceptual parallel between the residual and geometric curvature. The residual represents a failure of local connections (weights) to be compatible; the HPN evolution acts to "flatten" this local geometry, analogous to natural gradient flows.

In summary, the Hebbian Physics Network offers a novel, thermodynamically consistent architecture where physical laws emerge from local, adaptive interactions, removing the need for rigid scaffolding and global solvers in computational physics.