🔬 condensed matter

Graph neural network for multitask prediction of rheological and microstructural behavior in suspensions

This study proposes a microstructure-informed multitask graph neural network (GNN) that efficiently predicts the rheological properties and microstructural behavior of 2D suspensions by learning an implicit mapping from particle configurations, offering a computationally fast alternative to traditional simulations across various shear regimes.

Original authors: Armin Aminimajd, Joao Maia, Abhinendra Singh

Published 2026-02-10

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Armin Aminimajd, Joao Maia, Abhinendra Singh

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

The "Smart Traffic Controller" for Liquid Sand: A Simple Explanation

Imagine you are trying to drive a car through a massive, chaotic traffic jam. To predict how fast you’ll move, you usually have to look at every single car, calculate how close they are to each other, how much they are braking, and how much they are turning. In the world of science, this is called simulating a suspension (a liquid filled with tiny solid particles, like paint, mud, or even blood).

Doing this calculation for every single particle is incredibly slow and exhausting for even the most powerful supercomputers. It’s like trying to predict the movement of every single grain of sand in a desert by calculating the force of every tiny collision.

This paper introduces a new way to "cheat" the math using Artificial Intelligence.

The Problem: The "Chaos of the Crowd"

When you stir a liquid filled with particles, things get weird. If you stir slowly, it flows like water. But if you stir it hard, the particles suddenly crash into each other and lock up, making the liquid act like a solid. This is called "Shear Thickening" (think of how cornstarch and water turns into a solid when you punch it).

Predicting exactly when and how this "locking up" happens is a nightmare because it depends on the microscopic "handshakes" (contacts) between millions of tiny particles.

The Solution: The Graph Neural Network (The "Social Network" Approach)

Instead of using heavy physics equations to calculate every single force, the researchers built a Graph Neural Network (GNN).

The Analogy:
Think of the particles not as billiard balls, but as people at a crowded music festival.

Instead of calculating the physics of every person's muscles, the AI looks at the "Social Network" of the crowd.
It looks at who is standing near whom, how close the gaps are between people, and the general "shape" of the crowd.
By looking at the pattern of the crowd (the "graph"), the AI can instantly guess: "The crowd is getting too tight; a stampede (jamming) is about to happen!"

The "Multitask" Superpower

Most AI models are like specialists: one is good at math, one is good at art. This paper uses Multitask Learning, which is like a polymath—a person who is an expert in math, art, and history all at once.

Because the AI learns three things at the same time—how thick the liquid is (viscosity), how much pressure the particles exert (pressure), and how many particles are touching (coordination)—it actually becomes smarter. It realizes that if the "social network" of particles shows a lot of touching, the pressure must be high and the liquid must be thick. Learning these connections helps the AI correct its own mistakes.

Why Does This Matter?

The researchers found that their AI was incredibly accurate (99% accurate in many cases!). It could look at a single "snapshot" of where particles were located and instantly predict how the whole mixture would behave.

Real-world impact:

Industrial Efficiency: Companies making paint, food, or medicine can predict how their products will flow through pipes without running weeks of expensive computer simulations.
Medical Breakthroughs: It could help doctors better understand how blood (a complex suspension) flows through our veins.
Speed: It turns a "marathon" of calculation into a "sprint," allowing for real-time control of manufacturing processes.

In short: They taught a computer to "read the room" of microscopic particles to predict the behavior of the whole crowd, instantly.

Technical Summary: Graph Neural Network for Multitask Prediction of Rheological and Microstructural Behavior in Suspensions

1. Problem Statement

Predicting the rheological behavior of dense particulate suspensions is a significant computational challenge. Traditional methods, such as Lubrication Flow-Discrete Element Method (LF-DEM) simulations, are extremely computationally expensive because they must explicitly evaluate complex contact networks, hydrodynamic lubrication forces, and stress tensors, especially in regimes approaching shear thickening (ST) and jamming.

Furthermore, experimental characterization is often hindered by the inability to distinguish between particle pressure, fluid contribution, and frictional contacts due to phenomena like wall-slip or particle migration. There is a critical need for a "mesoscale surrogate"—a model that can rapidly infer macroscopic properties from microscopic structural snapshots without the heavy cost of full force-based simulations.

2. Methodology

The authors propose a microstructure-informed multitask learning (MTL) framework utilizing a Deep Graph Convolutional Neural Network (DeepGCN).

Data Generation: High-fidelity 2D suspension data were generated using LF-DEM simulations for semi-dilute to dense systems ( $\phi \in [0.70, 0.80]$ ) under controlled shear stress ( $\sigma/\sigma_0 \in [0.5, 200]$ ).
Graph Representation: The suspension configurations are treated as graphs where:
- Nodes: Represent particles, with features including particle radius (one-hot encoded for bidisperse systems).
- Edges: Represent interparticle interactions, with features including dimensionless gaps, relative distances ( $R_{ij}$ ), and trigonometric components ( $\sin/\cos$ ) of the relative positions to capture deformation direction.
Architecture (DeepGCN): To prevent the vanishing gradient problem common in deep networks, the model employs Residual Graph Convolutional layers (ResGCN). This allows the network to update node features through message passing (exchanging information between neighbors) while maintaining stable gradients via residual connections.
Multitask Learning (MTL): Instead of training separate models, a single network is trained to simultaneously predict three highly correlated properties:
1. Relative Viscosity ( $\eta_r$ )
2. Particle Pressure ( $\Pi$ )
3. Frictional Coordination Number ( $Z_\mu$ )
  The model uses a shared latent representation, which improves robustness and reduces overfitting by leveraging the inherent physical correlations between these targets.
Training Protocol: The researchers used a 5-fold cross-validation strategy and the Huber loss function to balance sensitivity to outliers with smoothness near zero.

3. Key Contributions

Minimalist Feature Input: The model achieves high accuracy using only geometric/topological features (positions and radii), deliberately excluding explicit knowledge of interparticle forces or processing parameters like shear stress.
Multitask Synergy: Demonstrates that predicting microstructural ( $Z_\mu$ ) and rheological ( $\eta_r, \Pi$ ) properties together improves the model's ability to capture the underlying physics of the suspension.
DeepGCN Implementation: Successfully applies deep residual graph architectures to solve the non-linear, fluctuating physics of near-jamming regimes.

4. Results

High Accuracy: For semi-dilute and dense systems ( $\phi \leq 0.76$ ), the model demonstrated exceptional performance with correlation coefficients $R^2 \approx 0.99$ and narrow mean absolute errors (MAE).
Handling Jamming Transitions: As the system approaches the frictional jamming limit ( $\phi \to \phi^\mu_J$ ), prediction scatter increases due to the inherent physical fluctuations and noise in the suspension (e.g., discontinuous shear thickening and shear jamming). However, even in these "noisy" regimes ( $\phi = 0.80$ ), the model remains in excellent agreement with the trends of the simulated flow curves.
System Size Robustness: Testing on both small ( $N=400$ ) and large ( $N=2000$ ) systems showed that while viscosity prediction is sensitive to finite-size effects, particle pressure and coordination number predictions are highly robust across different system sizes.
Trend Capture: The model successfully captures the non-monotonic "S-shaped" flow curves characteristic of Discontinuous Shear Thickening (DST).

5. Significance

This work provides a computationally efficient mesoscale surrogate for suspension rheology. By bypassing the need for explicit stress and force evaluations during inference, the model enables:

Accelerated Simulations: Rapidly exploring the phase space of particulate materials.
Real-time Control: Potential application in industrial settings for the real-time characterization and control of complex fluid flows.
Generalizability: The framework is applicable to a broader class of soft matter and network-based systems (like gels) where the relationship between structure and function is paramount.