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

arXiv:2304.06800 · cond-mat.mes-hall, cond-mat.mtrl-sci · Submitted 2023-04-13 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Quantum Radio. Generated commentary on the latest quantum physics and condensed matter papers.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "Predicting the Fracture Propensity of Amorphous Silica Using Molecular Dynamics Simulations and Machine Learning".

Mira: The gist Amorphous silica (a-SiO2) is a widely used inorganic material, and this study combines molecular dynamics simulations and machine learning to examine its molecular deformations and fracture mechanisms <ref:2304.06800#pg16>.

Kai: First, who's behind it and why it matters.

Title and authors: Kai: So we're looking at this paper now, it's called "Predicting the Fracture Propensity of Amorphous Silica Using Molecular Dynamics Simulations and Machine Learning." It sounds pretty technical, but basically they’re using molecular dynamics simulations and machine learning to figure out how amorphous silica breaks.

Mira: Yeah, it tackles something that’s really important because amorphous silica is used everywhere—semiconductor chips, window panels—but it tends to be brittle when things get stressed. This paper seems to want to look at the atomic level deformations and fracture mechanisms which are usually hard to see directly.

Lev: It's interesting that they're using those two big tools together, MD simulations and machine learning. From where I sit, we always worry about how well those simulations actually capture the real physics of things like bond switching or fracture initiation.

Kai: Exactly, so the authors are trying to connect the microscopic atomic arrangements—like coordination numbers—to how ductile or brittle a sample will be under stress. It suggests that understanding these local structures is key to predicting material failure.

Mira: And they seem to find that by quenching at higher pressures, which makes the silica denser, you actually increase the ductility and toughness of the material, even though it might not have stronger bulk properties at those higher pressures.

Lev: That's a crucial point for real hardware design because it tells us how structural densification can trade strength for energy absorption capability.

Kai: Right, so we're seeing a complex relationship where density and ductility go in different directions, which is something to keep in mind when designing these materials.

The paper's summary: Mira: Looking at the summary of "Predicting the Fracture Propensity of Amorphous Silica Using Molecular Dynamics Simulations and Machine Learning," they are focusing on how local bonding topologies facilitate energy dissipation during plastic deformation, especially when the coordination numbers decrease.

Kai: They're specifically looking at three types of changes in the local atomic environment during tension: a decrease in coordination number, which they call DCN, an increase in coordination number, ICN, and no change at all. They say that the numbers for DCN and ICN atoms generally go up together as you apply more strain.

Lev: That suggests that bond-switching events are happening more frequently as the material deforms under load. We'd have to check if those frequency changes translate to measurable energy dissipation in a real system.

Mira: The paper then links this DCN activity directly to fracture propensity by constructing a classification-based machine learning model using three different algorithms. This model is designed to predict the bond-switching activities upon fracture based on the initial structure.

Kai: So, they're not just simulating the deformation; they're training an AI to look at what you start with and predict how likely that material is to fracture in a certain way. That’s where it gets really interesting for understanding the atomistic origins of failure.

Lev: If that machine learning model can actually predict the final state from the initial configuration, it means we could potentially use that knowledge to guide material synthesis toward more ductile structures.

Mira: It seems like they've managed to bridge the gap between atomic structure and macroscopic mechanical response by using this data-driven approach.

The paper's improvements: Kai: One of the main things they highlight as an improvement is linking the atomic strain metric to crack position. They found that fracture only happens at locations with maximum atomic strain compared to other spots in the solid.

Mira: And they made a quantitative connection by showing that atoms with high atomic strains and atoms with decreased coordination numbers largely occupy the same regions, which they say is especially true near where the fracture surface ends up.

Lev: That spatial correlation between DCN and strain is a strong piece of evidence because it moves beyond just saying "DCN matters" to showing *where* it matters structurally.

Kai: And they even predicted specific locations for fracture initiation using this correlation, finding positions at zero point seven seven eight and zero point seven seven two, which suggests a very precise prediction based on the initial configuration of the sample.

Mira: They also used machine learning to predict changes in coordination number upon fracture by training two binary classifiers, DCN-ICN and DCN-UCN, and both achieved prediction accuracies around ninety percent.

Lev: Ninety percent accuracy is respectable for a complex mechanical process like bond switching, especially when trying to distinguish between those different states from the initial input structure.

Kai: So the improvement here is really about establishing this connection: you can look at an unstrained amorphous silica sample and use its initial atomic configuration to predict its fracture propensity very well using these AI classifiers.

Conclusion: Mira: To wrap up "Predicting the Fracture Propensity of Amorphous Silica Using Molecular Dynamics Simulations and Machine Learning," the authors conclude that the development of decreased coordination number activities is closely tied to the deformation mechanisms, and they pinpoint DCN as a key factor in learning how propensity to fracture.

Kai: They successfully established a connection between initial structure, DCN, and the position of fracture within the material. Basically, you can predict how likely a silica sample is to fail based on its starting atomic configuration before any stress is applied.

Lev: From an engineering standpoint, this means we could potentially use that initial atomistic configuration to screen materials for better ductility before even synthesizing them in bulk quantities.

Mira: It’s also interesting because the DCN-ICN and DCN-UCN classifiers both achieved around ninety percent prediction accuracy when predicting the final state based on the initial environment.

Kai: That means we can successfully distinguish those different states of coordination number at fracture just by looking at what the atoms were like when they started out. It’s a really neat way to connect structure to mechanical behavior.

Lev: For real hardware, this suggests a pathway toward designing materials where we can have high confidence in their mechanical response based on initial design parameters rather than just testing them after they're made.

Kai: So that's the gist of "Predicting the Fracture Propensity of Amorphous Silica Using Molecular Dynamics Simulations and Machine Learning." It shows how combining MD and ML can give us deep insights into amorphous solids, even when we’re dealing with complex fracture mechanisms.

JIAHAO LIU, JINGJIE YEO

Sibley School of Mechanical and Aerospace Engineering, Cornell University · Department of Engineering Mechanics, Xi’an Jiaotong University

cond-mat.mes-hall, cond-mat.mtrl-sci

Submitted: 2023-04-13

Updated: 2023-04-13

Comments: 26 pages, 13 figures

DOI: 10.1142/S1758825123500862

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 79/100

The gist: The gist Amorphous silica (a-SiO2) is a widely used inorganic material, and this study combines molecular dynamics simulations and machine learning to examine its molecular deformations and fracture

Key concepts

Coordination Number (CN): The CN measures how many neighboring atoms surround a specific atom in the silica structure. Changes in this number—increasing (ICN), decreasing (DCN), or staying the same (UCN)—are monitored during deformation. DCN is identified as a critical factor linking atomic arrangement to fracture likelihood.
Fracture Propensity Prediction: This refers to the ability of an amorphous silica sample to break under stress. The researchers used machine learning classifiers trained on initial atomic features (like structure functions) to predict whether a sample would fracture, specifically focusing on predicting the final state's DCN based on its starting structure.
Atomic Strain and Crack Position: Atomic strain measures how much the local atomic arrangement is distorted during deformation. The study found that atoms experiencing high atomic strain strongly correlate with atoms exhibiting a decreased coordination number (DCN), suggesting that fracture initiation occurs precisely where these two features overlap.
Brittle to Ductile Transition: This describes the change in material behavior observed when silica is deformed under tension. Increasing the quenching pressure caused a transition where samples became more ductile, meaning they could sustain greater plastic deformation before fracturing, leading to higher toughness.

Summary

The gist Amorphous silica (a-SiO2) is a widely used inorganic material, and this study combines molecular dynamics simulations and machine learning to examine its molecular deformations and fracture mechanisms <ref:2304.06800#pg16>. This research is essential for understanding how atomic arrangements strongly influence the mechanical properties and structural features in amorphous solids <ref:2304.06800#pg20>.

Mechanical Response in Macroscale

The study examined the mechanical response of a-SiO2 samples deformed under uniaxial tension using MD simulations with LAMMPS <ref:2304.06800#pg22>. A clear brittle to ductile transition was observed as the quenching pressure increased, indicating that densification increases ductility. Specifically, a sample quenched at 5 GPa might not fracture until > 40% uniaxial strain was attained, doubling that of lower-pressure quenched samples. The toughness for different pressure-quenched samples increased monotonically as the quenching pressure rose from 0 to 5 GPa, reaching a maximum value of 4.456 GJ/m3. This suggests that even though more densified a-SiO2 at higher quenching pressures had poorer mechanical strengths, the enhanced ductility dominated the energy absorption process and led to higher toughness for the material.

Microscale Bond-Switching Events

The researchers monitored bond-switching events by measuring coordination numbers (CN) during tensile deformation. The local atomic environment could change in three ways: a decrease in coordination number (DCN), an increase in coordination number (ICN), or the coordination number remained unchanged (UCN). The numbers of DCN and ICN atoms generally increased in tandem with the applied strain, suggesting that more and more bond-switching events occurred as the aSiO2 samples were progressively deformed. More densified samples had higher numbers of bond-switching events throughout the deformation process as well. The average number of DCN atoms upon fracture grew significantly as the quenching pressure increased, which helps rationalize why samples quenched under higher pressure had better nano-ductility.

Linking CNC with Crack Position

The researchers used atomic strain as a metric for fracture, noting that crack only occurs at locations with maximum atomic strain compared to other locations in the solid. They found that atoms with high atomic strains and atoms with decreased CN largely resided in the same regions, which was extremely pronounced in the vicinity of the eventual fracture surface. Quantitatively, the distribution functions for DCN atoms and atomic strain mostly overlapped in the same region for samples quenched at 0 GPa. This demonstrated a strong correlation between DCN and atomic strain, where the predicted positions of fracture initiation were 0.778 and 0.772 respectively.

Linking Initial Atomic Environment with CNC

The study used machine learning to predict changes in CN upon fracture by training a classifier on initial atomic configurations. The input features included 47 radial structure functions (RSFs) and 22 angular structure functions (ASFs), totaling 69 input features. The researchers trained two binary ML classifiers, DCN-ICN and DCN-UCN, to predict the change in CN from the initial state to the final state. Both classifiers achieved prediction accuracies of around 90% for both DCN-ICN and DCN-UCN classifiers, implying that the DCN at the final state could be successfully distinguished based on the initial atomic environment. This established a connection between initial structure-DCN-fracture position and predicted fracture propensity.

Conclusion

The study concluded that the development of decreased CN activities was closely related to the deformation mechanisms and inferred that DCN was the key factor in learning the fracture propensity. The connection of initial structure-DCN-fracture position was established and the fracture propensity could be predicted based on the initial atomistic configuration of an unstrained a-SiO2 sample. This suggests that both DCN-ICN classifier and DCN-UCN classifier could achieve a prediction accuracy of around 90%, which implied that the DCN at the final state could be successfully distinguished based on the initial atomic environment.

Methods

The amorphous phase construction involved melting silica and randomizing its structure through a melt-quenching process under various pressures. Fracture simulations were performed under uniaxial tensile deformations, focusing on the “plane stress” condition with a strain rate of 10 9 s-1. The coordination number was determined by counting neighboring atoms within a set cutoff distance of 1.8 Å. Structure functions, including RSFs and ASFs, were used to generate the input vector Xi with a total number of 69 input features. The ML methodology involved training classifiers using Nearest Neighbors, Support Vector Machines (Linear SVM, Polynomial SVM, RBF SVM), and Trees and Forest (Decision Tree, Extra Trees, Random Forest). The data preparation included min-max feature scaling to handle the range of values in the features.

References

1 Application of Silica Glasses in Practice. [1991]. In Glass Science and Technology (Vol. 11, pp. 271–300). https://doi.org/10.1016/B978-0-444-98755-6.50011-2

2 Argon, A. S. [1979]. Plastic deformation in metallic glasses. Acta Metallurgica, 27(1), 47–58. https://doi.org/10.1016/0001-6160(79)90055-5

3 Askeland, D. R., and Wright, W. J. [2016]. The science and engineering of materials (Seventh edition). Boston, MA: Cengage Learning. https://doi.org/10.1179/1743280410Y.0000000

4 Behler, J., and Parrinello, M. [2007]. Generalized Neural-Network Representation of HighDimensional Potential-Energy Surfaces. Physical Review Letters, 98(14), 146401. https://doi.org/10.1103/PhysRevLett.98.146401

5 Breiman, L. [2001]. Random Forests. Machine Learning, 45(1), 5–32. https://doi.org/10.1023/A:1010933404324

6 Butler, K. T., Davies, D. W., Cartwright, H., Isayev, O., and Walsh, A. [2018]. Machine learning for molecular and materials science. Nature, 559(7715), 547–555. https://doi.org/10.1038/s41586-018-0337-2

7 Cao, Y., Li, J., Kou, B., Xia, C., Li, Z., Chen, R., … Wang, Y. [2018]. Structural and topological nature of plasticity in sheared granular materials. Nature Communications, 9(1), 2911. https://doi.org/10.1038/s41467-018-05329-8

8 Chang, C.-C., and Lin, C.-J. [2011].

Improvements for AI systems

Improved Fracture Propensity Prediction Accuracy via ML Classifiers: The improved AI system can distinguish between DCN states with high accuracy, as both DCN-ICN classifier and DCN-UCN classifier could achieve a prediction accuracy of around 90% when predicting the final state from the initial configuration, implying that the DCN at the final state could be successfully distinguished based on the initial atomic environment.
Enhanced Fracture Position Localization: The AI system can predict crack initiation locations with high precision by correlating structural features, as the distribution functions for DCN atoms and atomic strain were mostly overlapped in the same region, quantitively demonstrating the strong correlation between DCN and atomic strain, leading to predicted fracture positions being almost identical to exact positions.
Material Property Mapping via Pressure Quenching: The improved system can predict macroscopic mechanical properties based on initial structure, as it can demonstrate that more densified samples were always associated with higher failure strains and toughnesses, enabling the prediction of material behavior from the quenching pressure alone.

Abstract

Amorphous silica (a-SiO 2) is a widely used inorganic material. Interestingly, the relationship between the local atomic structures of a-SiO 2 and their effects on ductility and fracture is seldom explored. Here, we combine large-scale molecular dynamics simulations and machine learning methods to examine the molecular deformations and fracture mechanisms of a-SiO 2. By quenching at high pressures, we demonstrate that densifying a-SiO 2 increases the ductility and toughness. Through theoretical analysis and simulation results, we find that changes in local bonding topologies greatly facilitate energy dissipation during plastic deformation, particularly if the coordination numbers decrease. The appearance of fracture can then be accurately located based on the spatial distribution of the atoms. We further observe that the static unstrained structure encodes the propensity for local atomic coordination to change during applied strain, hence a distinct connection can be made between the initial atomic configurations before loading and the final far-from-equilibrium atomic configurations upon fracture. These results are essential for understanding how atomic arrangements strongly influence the mechanical properties and structural features in amorphous solids and will be useful in atomistic design of functional materials.