🔬 condensed matter

Using graph neural networks to predict many-body interactions in amorphous materials

This paper demonstrates that the equivariant graph neural network NequIP can accurately and efficiently predict many-body interactions in solvent-free polymer-grafted nanoparticles by learning from classical density functional theory, enabling the discovery of experimentally consistent equilibrium structures despite being trained solely on out-of-equilibrium configurations.

Original authors: Mehryar Jannesari Ghomsheh, Donald L. Koch, Sarah Hormozi

Published 2026-06-01

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Mehryar Jannesari Ghomsheh, Donald L. Koch, Sarah Hormozi

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 a crowd of people standing in a room, each wearing a giant, fluffy, elastic coat. These coats are so big that they fill every single inch of empty space in the room. Now, imagine trying to move one person. Because their coat is tangled with everyone else's, you can't just step aside; you have to push, pull, and rearrange the coats of your neighbors to make room. This is a bit like what happens in a special type of material called "solvent-free polymer-grafted nanoparticles" (PGNs).

In these materials, tiny solid cores (the people) are covered in long polymer chains (the coats). When there is no liquid solvent between them, these chains are forced to stretch out and fill every gap, creating a tight, interconnected web. Moving one particle requires a complex, coordinated dance involving many neighbors at once. Scientists call these "many-body interactions," and they are notoriously difficult to calculate because they depend on the exact shape and angle of the crowd, not just simple distance.

The Problem: A Rugged Mountain Range
To understand how these materials behave, scientists need to map their "Potential Energy Landscape" (PEL). Think of this landscape as a giant, rugged mountain range.

High-energy spots are like steep, jagged peaks where the particles are squeezed uncomfortably (like the coats being stretched too tight).
Low-energy spots are deep, smooth valleys where the particles are happy and relaxed.

Traditionally, calculating the height of every point on this mountain range is like trying to climb every single hill in the world. It takes so much computer power that scientists usually have to simplify the problem, pretending the particles only interact with their immediate neighbors (like a simple handshake). But in these soft materials, that simplification misses the whole story because the "coats" create complex, multi-person interactions.

The Solution: A Smart AI Guide
The researchers in this paper used a type of Artificial Intelligence called a Graph Neural Network (GNN), specifically a model named NequIP, to solve this.

Think of NequIP as a super-smart guide who has been shown thousands of photos of the mountain range in its most chaotic, high-energy states (the jagged peaks). The guide learns the rules of the terrain: "If the coats are stretched this way, the energy goes up; if they are tangled that way, it goes down."

Crucially, this AI doesn't just look at how far apart two particles are. It understands the geometry of the crowd. It knows that if three particles form a triangle, the energy is different than if they form a line. It learns the "shape" of the interactions.

The Magic Trick: Predicting the Peaceful Valley
Here is the most surprising part of the paper: The AI was only trained on the chaotic, high-energy, "out-of-equilibrium" states (the jagged peaks). It never saw the peaceful, low-energy valleys where the material naturally settles.

However, when the researchers let the AI drive a simulation to find the bottom of the valley (the equilibrium state), it succeeded.

The Result: The AI found the exact same peaceful structures that real experiments observe.
The Analogy: It's like showing a hiker only pictures of a stormy, rocky mountain peak and asking them to find the calm, sunny meadow at the bottom. Even though they never saw the meadow, they understood the rules of the mountain so well that they could navigate straight to it.

What Did They Discover?
By using this AI, the researchers found that these materials naturally form tiny, hidden clusters that look like icosahedrons (shapes with 20 triangular faces, like a soccer ball).

These shapes are special because they pack together very tightly, allowing the polymer "coats" to stretch out evenly without getting tangled.
The AI revealed that these structures only appear when the material is in its relaxed, equilibrium state, not when it's being jostled around.

Why This Matters
The paper shows that we can use this AI to predict how these complex, soft materials behave with incredible accuracy, but at a cost that is 10,000 times cheaper than traditional methods.

Instead of spending years calculating every possible interaction, the AI learns the "rules of the game" from a few thousand examples and then plays the game instantly.
This allows scientists to explore the deep, hidden valleys of the energy landscape that were previously too expensive to reach, revealing the true, natural structure of these soft, glassy materials.

In short, the paper demonstrates that a smart AI, trained on chaos, can perfectly predict order, helping us understand how complex soft materials settle down into their most stable forms.

Technical Summary: Using Graph Neural Networks to Predict Many-Body Interactions in Amorphous Materials

Problem Statement
Many-body interactions are fundamental to the behavior of amorphous materials, ranging from metallic glasses to biological tissues. However, computational modeling of these systems has historically been restricted to pairwise-additive potentials due to the prohibitive cost of accurately resolving higher-order interactions. This limitation prevents a full characterization of the high-dimensional, rugged potential energy landscape (PEL) that governs the slow dynamics and glassy behavior of these materials. While machine learning interatomic potentials (MLIPs) have shown promise in modeling atomic interactions, their ability to resolve the complex PEL of soft, many-body glassy systems—specifically those dominated by angular-dependent interactions—remains an open question. Furthermore, existing MLIPs have not been extended to soft glasses, and it is unclear if they can generalize from high-energy, out-of-equilibrium training data to low-energy equilibrium states.

Methodology
The authors address this challenge by applying the equivariant message-passing graph neural network (GNN), NequIP, to a model soft glass consisting of solvent-free polymer-grafted nanoparticles (PGNs). In this system, the absence of solvent forces grafted polymer chains to uniformly fill the interstitial space, creating strong, angular-dependent many-body interactions between particle cores that cannot be captured by pairwise potentials.

The workflow involves three primary stages:

Data Generation via Classical DFT: The authors utilize classical density functional theory (DFT) to model polymer-mediated interactions. They sample 5,000 uncorrelated, high-energy, out-of-equilibrium configurations of PGNs using hard-sphere dynamics. The potential energy for each configuration is calculated using classical DFT, which treats the polymers as an incompressible fluid filling the void space. This generates a dataset spanning the high-energy region of the PEL.
GNN Training: NequIP is trained on this dataset to learn the mapping from particle coordinates to potential energy. The network utilizes tensor representations that are equivariant to translations, rotations, and reflections, allowing it to capture anisotropic many-body effects. The authors systematically analyze hyperparameters, including the cutoff radius ( $r_{cut}$ ), the number of message-passing layers ( $n_l$ ), and the maximum angular resolution ( $l_{max}$ ), to determine the physical range and order of interactions.
GNN-Driven Monte Carlo (MC) Simulations: Once trained, the GNN replaces the classical DFT in Monte Carlo simulations. This acceleration (four orders of magnitude faster) enables the exploration of the deep minima of the PEL, allowing the system to reach equilibrium configurations that were inaccessible during the training phase.

Key Contributions and Results

Accurate PEL Learning: The trained NequIP model reproduces classical DFT energies with a mean absolute error (MAE) of approximately 0.01 $\sigma_E$ (standard deviation of energy) across various design parameters (grafting density, core volume fraction, polymer molecular weight). This corresponds to per-particle errors well below the thermal energy scale ( $k_B T$ ).
Physical Insights into Interaction Range and Order: Systematic analysis of hyperparameters reveals that accurate predictions require a large cutoff radius ( $r_{cut}/d \approx 6$ ), indicating that effective interactions extend well beyond immediate neighbors due to many-body correlations rather than long-range pair forces. Furthermore, the error convergence with respect to the number of message-passing layers ( $n_l$ ) suggests that four-body correlations are crucial for energy prediction in solvent-free PGNs, contrasting with solvent systems where three-body interactions often suffice. The necessity of equivariant representations ( $l_{max} \ge 1$ ) confirms the highly directional nature of these interactions.
Out-of-Distribution Generalization: A striking result is the network's ability to predict equilibrium structures despite being trained exclusively on high-energy, out-of-equilibrium configurations. GNN-driven MC simulations recover equilibrium structures that agree with experimental measurements (e.g., nearest-neighbor distances from SAXS) and classical DFT predictions. The equilibrium energy distribution lies well outside the training distribution, yet the GNN maintains DFT-level accuracy, demonstrating robust generalization.
Discovery of Local Order: Analysis of the equilibrium structures reveals the absence of crystal-like ordering but the presence of locally favored icosahedral-like structures. These structures are strongly correlated with specific Voronoi volumes (dodecahedral shapes) that minimize the free energy of the grafted polymers. This local ordering is observed only at equilibrium and is absent in the out-of-equilibrium training data.
Hyperuniformity: The equilibrium structures exhibit suppressed long-range density fluctuations (hyperuniformity), a feature consistent with experimental observations of solvent-free PGNs, which is not present in the out-of-equilibrium hard-sphere sampled configurations.

Significance
The paper demonstrates that equivariant GNNs can successfully learn the high-dimensional, rugged potential energy landscapes of soft glassy systems dominated by many-body interactions. By training on computationally accessible high-energy data, the model can predict equilibrium properties and structures that are otherwise computationally prohibitive to obtain via classical DFT. This work establishes a practical workflow for studying amorphous materials where many-body effects are non-negligible, offering a path toward inverse design and a deeper understanding of the structure-dynamics relationship in glassy systems. The findings suggest that hidden structural features shared between equilibrium and non-equilibrium states can be extracted by ML models, potentially motivating new physics-based order parameters.

Technical Summary: Using Graph Neural Networks to Predict Many-Body Interactions in Amorphous Materials

More like this