🔬 mesoscale physics

Predicting the Fracture Propensity of Amorphous Silica Using Molecular Dynamics Simulations and Machine Learning

This study combines molecular dynamics simulations and machine learning to reveal that high-pressure densification enhances the ductility and toughness of amorphous silica by altering local bonding topologies, thereby establishing a predictive link between initial atomic configurations and fracture propensity.

Original authors: Jiahao Liu, Jingjie Yeo

Published 2026-10-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jiahao Liu, Jingjie Yeo

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

Glass is everywhere, from the windows of our homes to the screens of our phones, yet it remains one of the most misunderstood materials in engineering. At the heart of this mystery is a simple question: why does glass shatter so easily? Unlike metals, which can bend and stretch before breaking, glass usually snaps suddenly and catastrophically. This brittleness is a major limitation for engineers who want to build stronger, safer structures. For decades, scientists have known that the internal arrangement of atoms in glass is different from the orderly crystals found in metals or diamonds, but they have struggled to connect that invisible atomic mess to the visible way the material fails. The key to unlocking this puzzle lies in understanding how the tiny building blocks of glass rearrange themselves when pulled or squeezed, and whether those rearrangements can be predicted before the material ever breaks.

In a recent study, researchers set out to solve this problem by looking at amorphous silica, the primary ingredient in most common glass. They wanted to see if they could predict exactly where a piece of glass would crack just by looking at its structure before any force was applied. To do this, they used powerful computer simulations to create virtual samples of silica and then pulled them apart, mimicking a tensile test. They discovered that the way they made the glass mattered immensely. By cooling the molten silica under high pressure, they created denser versions of the material that behaved very differently from standard glass. While normal glass would snap after stretching just a tiny bit, these high-pressure samples could stretch nearly three times as far before failing, transforming from a brittle material into one that could bend and deform without shattering.

The secret to this newfound flexibility was found in the way the atoms moved. In standard glass, the atoms are locked in place, and when the material is pulled, the stress builds up until it suddenly releases in a crack. In the denser, high-pressure glass, the atoms had more room to shift. As the material stretched, the bonds between atoms would break and reform in new configurations, a process the researchers called bond-switching. This constant reorganization allowed the material to absorb energy and spread out the stress, preventing a single crack from forming immediately. The researchers found that the most important sign of an impending break was a specific type of atomic change: when an atom lost a neighbor it was bonded to. By tracking these moments of lost connections, they could map out exactly where the material was weakening long before the crack actually appeared.

To turn this observation into a predictive tool, the team employed a technique known as machine learning. They taught a computer program to recognize the unique atomic patterns present in the glass before it was stretched. The computer learned to look at the initial arrangement of atoms and predict which ones were likely to lose their connections when the material was eventually pulled. Remarkably, the computer was able to identify the future location of the fracture with high accuracy, simply by analyzing the static, unstrained structure. It did not need to see the glass deforming; the blueprint for the break was already encoded in the starting position of the atoms. This suggests that the potential for a material to fail is not random but is determined by its initial atomic architecture.

The implications of this work extend beyond just understanding glass. The study demonstrates that the mechanical properties of a material, such as its ability to bend or its tendency to shatter, are deeply rooted in its microscopic structure. By using computer simulations to test different ways of making glass, the researchers showed that it is possible to engineer materials that are tougher and more resilient. The ability to predict where a crack will start based on the initial structure opens the door to designing new materials that can withstand extreme conditions. Instead of guessing which designs will work, engineers could potentially use these methods to screen atomic arrangements and select only those that promise the desired strength and flexibility, leading to a new generation of durable materials for everything from aerospace to everyday consumer goods.

Technical Summary: Predicting the Fracture Propensity of Amorphous Silica Using Molecular Dynamics Simulations and Machine Learning

Problem Statement
Amorphous silica (a-SiO₂) is a critical material for applications ranging from semiconductor chips to optical fibers. However, it is inherently susceptible to brittle fracture, particularly in the presence of surface flaws, which limits its engineering utility. While the deformation mechanisms of crystalline solids are well-understood, the origins of ductility and fracture propensity in non-crystalline materials like a-SiO₂ remain challenging to determine. Previous studies have proposed shear transformation zones (STZs) as carriers of ductility, but these are more commonly studied in metallic glasses, and the brittle nature of oxide glasses obscures the existence of such plastically deforming regions. Furthermore, the relationship between the local atomic structures of a-SiO₂ and their specific effects on ductility and fracture has been seldom explored.

Methodology
The authors employed a combined approach of large-scale molecular dynamics (MD) simulations and machine learning (ML) to investigate the deformation and fracture mechanisms of a-SiO₂.

MD Simulations:
- Sample Preparation: A series of a-SiO₂ samples were generated using a melt-quenching process. Starting from a crystalline block, the system was melted at 5000 K and then cooled to 300 K under various constant hydrostatic pressures (0, 1, 2, 3, 4, and 5 GPa) to create samples with varying degrees of densification.
- Deformation: Both fully periodic bulk samples and nanowires with free surfaces were subjected to uniaxial tensile deformation at a strain rate of $10^9 s^{-1}$ at 300 K.
- Analysis: The authors utilized the Open Visualization Tool (OVITO) and Visual Molecular Dynamics (VMD) to analyze atomic configurations. Key metrics included stress-strain curves, atomic strain tensors (volumetric and shear), and coordination numbers (CN).
- Bond-Switching Events: Atoms were categorized based on changes in their coordination numbers (CN) between the initial unstrained state and the final fractured state: Decreased CN (DCN), Increased CN (ICN), and Unchanged CN (UCN).
Machine Learning:
- Feature Engineering: The initial static atomic configurations were described using 69 structural descriptors: 47 Radial Structure Functions (RSFs) and 22 Angular Structure Functions (ASFs).
- Classification Task: The goal was to predict the bond-switching activity (DCN, ICN, or UCN) of an atom in the final fractured state based solely on its initial unstrained environment.
- Algorithms: Three major classes of supervised learning algorithms were tested: Nearest Neighbors (KNN), Support Vector Machines (Linear, Polynomial, and RBF kernels), and Tree-based methods (Decision Tree, Random Forest, Extra Trees).
- Data Handling: The dataset was balanced (350 atoms per class) and normalized using min-max scaling. Models were trained on 70% of the data and tested on 30%, with hyper-parameters tuned via grid search.

Key Results

Macroscopic Mechanical Response:
- Increasing the quenching pressure induced a clear brittle-to-ductile transition. Samples quenched at low pressures (< 2 GPa) exhibited sudden brittle failure at strains < 0.2.
- Samples quenched at higher pressures (up to 5 GPa) showed significantly increased ductility, with failure strains reaching > 0.58 (nearly double that of low-pressure samples).
- While higher pressure increased ductility and toughness (energy absorption per unit volume), it decreased the ultimate tensile strength (maximum stress dropped from ~13 GPa to ~9 GPa).
- Nanowires exhibited similar trends but with lower overall strength and ductility compared to bulk periodic samples, attributed to free surface effects and necking.
Microscopic Mechanisms:
- Bond-Switching: The number of bond-switching events (both DCN and ICN) increased with applied strain and was significantly higher in densified samples.
- Role of DCN: The authors identified Decreased Coordination Number (DCN) events as the primary indicator of fracture propensity. The number of DCN atoms grew significantly with pressure and correlated strongly with the spatial location of crack initiation.
- Fracture Localization: The spatial distribution of DCN atoms overlapped almost perfectly with regions of maximum atomic strain. The peak of the DCN distribution function accurately predicted the precise position of crack initiation (e.g., predicted at 0.778 vs. exact at 0.772).
Machine Learning Performance:
- The ML classifiers successfully mapped the initial static atomic structure to the future bond-switching state.
- Four out of seven tested algorithms achieved prediction accuracies of 90% or higher.
- The RBF SVM and Random Forest models performed particularly well, with the DCN-ICN classifier reaching up to 92.57% accuracy and the DCN-UCN classifier reaching 91.43%.
- Data normalization significantly improved the performance of distance-based algorithms (KNN, SVM), while tree-based algorithms were less sensitive to normalization.
- Including Angular Structure Functions (ASFs) provided a marginal improvement over using only Radial Structure Functions (RSFs), suggesting RSFs contain most of the necessary local structural information.

Significance and Claims
The paper claims to establish a distinct connection between the initial static atomic configuration of amorphous silica and its final far-from-equilibrium state upon fracture. The primary significance lies in demonstrating that:

Ductility is Tunable: Densifying a-SiO₂ via high-pressure quenching enhances toughness and ductility by facilitating local bond-switching events that dissipate energy.
Predictive Capability: The static, unstrained structure encodes the propensity for local atomic coordination changes during deformation.
Fracture Prediction: It is possible to accurately predict the location of fracture and the specific bond-switching activities (specifically DCN) using only the initial atomic configuration and machine learning, without simulating the deformation process.

The authors conclude that these findings provide a holistic insight into how atomic arrangements influence mechanical properties in amorphous solids. They suggest this approach is useful for the atomistic design of functional materials, allowing for the prediction of fracture propensity based on initial structural features. The study does not claim to solve all fracture problems in oxide glasses but provides a validated framework for linking static structure to dynamic failure in a-SiO₂.

More like this