Modeling phase separation in polymer-derived silicon carbonitride ceramics through extended machine learning molecular dynamics

summary

The gist

Modeling phase separation in polymer-derived silicon carbonitride ceramics through extended machine learning molecular dynamics develops and applies a machine learning interatomic potential for…

In short

Researchers developed a machine learning interatomic potential for silicon carbonitride (SiCN) systems to simulate phase separation during thermal processing. The model successfully predicted the formation of distinct graphene-like carbon sheets nucleating from an amorphous SiCN matrix, matching experimental atomic pair distribution functions and revealing the material's unique ceramic and graphitic properties.

Key concepts

Machine Learning Interatomic Potential (MLIP)
This is a computational tool trained on vast amounts of structural data to predict the forces and energies of atomic interactions in materials like SiCN. It allows for large-scale molecular dynamics simulations, enabling researchers to study how materials change structurally during high-temperature processes without needing computationally expensive traditional methods.
Phase Separation
This is a process where a material separates into two or more distinct phases with different compositions or structures. In this SiCN ceramic, the amorphous silicon-based network separates from a 'free carbon' phase as the material is heated, leading to the formation of separate carbon domains.
Pair Distribution Function (PDF)
The PDF is a mathematical tool used to characterize the local atomic structure of a material at different length scales. By comparing simulated PDFs with experimental ones, researchers can verify if their model accurately captures the real-world arrangement of atoms, especially in the medium-range order.
Free Carbon Phase
This is a specific structural feature within polymer-derived SiCN materials that forms during pyrolysis. It consists of disordered sp2-hybridized carbon clusters that organize into turbostratic phases above 1200°C. This phase is crucial because its metallic-like behavior significantly enhances the material's electrical conductivity.

Terminology used across episodes

This episode discusses

The paper

Modeling phase separation in polymer-derived silicon carbonitride ceramics through extended machine learning molecular dynamics · Read on arXiv

Institut de Recherche sur les Céramiques (IRCER) · Université de Limoges University of Limoges University of Strasbourg University of Strasbourg CNRS consortium ADYNMAT CNRS consortium Centre for Catalysis and Clean Energy Griffith University

Polymer-derived ceramics combine the thermal stability of ceramics with the versatile properties of carbon domains, but modeling their atomic-scale evolution during processing remains elusive due to the limitations of traditional computational methods. To address this issue, here we develop and apply a machine learning interatomic potential for silicon carbonitride-based (Si-C-N-H) systems, trained on a diversified database of over 9000 configurations -- including amorphous models, high-temperature states, surfaces, and crystal structure predictions -- to capture the full complexity of these materials. This potential enables large-scale molecular dynamics simulations of 8000-atom systems revealing the atomic-scale evolution of the polymer-derived ceramic during thermal treatment. A key result of this work is the occurrence of a phase separation where carbon domains progressively nucleate from the amorphous SiCN matrix during thermal processing, forming distinct graphene-like sheets while preserving the integrity of the ceramic network. The resulting models reproduce the experimental atomic pair distribution functions with exceptional fidelity, validating our approach and providing microscopic explanations for the material unique combination of ceramic and graphitic properties. In this process, defective 5- and/or 7-member carbon rings, mediate the transformation to stable 6-member aromatic structures. These findings offer new atomic-scale insights into the thermal stability and structural transformation pathways of polymer-derived ceramics, while our methodology opens avenues for studying complex amorphous systems with first-principles accuracy at experimentally relevant scales.

DOI: 10.1021/acsnano.6c11958

Transcript

Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.

Kai: I'm Kai.

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

Kai: We're looking at the paper: "Modeling phase separation in polymer-derived silicon carbonitride ceramics through extended machine learning molecular dynamics".

Mira: Here's the gist of the paper. Modeling phase separation in polymer-derived silicon carbonitride ceramics through extended machine learning molecular dynamics develops and applies a machine learning interatomic potential for silicon carbonitridebased (Si-C-N-H) systems, trained on a diversified database of over 9000 configurations to capture the full complexity of these materials.

Kai: A key result of this work is the occurrence of a phase separation where carbon domains progressively nucleate from the amorphous SiCN matrix during thermal processing, forming distinct graphene-like sheets while preserving the integrity of the ceramic network.

Mira: The resulting models reproduce the experimental atomic pair distribution functions with exceptional fidelity, validating our approach and providing microscopic explanations for the material unique combination of ceramic and graphitic properties.

Kai: Material Context Polymer-derived ceramics (PDCs) are advanced materials synthesized through pyrolysis of preceramic polymers such as polysilazanes, inorganic polysiloxanes or polycarbosilanes.

Title and authors: Kai: So we're talking about this paper now. It’s called "Modeling phase separation in polymer-derived silicon carbonitride ceramics through extended machine learning molecular dynamics." This work is about using machine learning to figure out how these complex materials separate into different phases when they get heated up.

Mira: Right, it deals with those polymer-derived ceramics, or PDCs, where you have this mixture of an amorphous silicon network and a "free carbon" phase that forms during processing. The authors are developing a machine learning interatomic potential for silicon carbonitride based systems, Si-C-N-H.

Lev: From an engineering standpoint, that sounds like they’re trying to bridge the gap between what we can simulate at the atomic level and what’s actually happening in the furnace. It's about capturing this phase separation during thermal processing.

Kai: Exactly, and they trained this machine learning potential on a big database of over nine thousand configurations to get a full picture of these Si-C-N-H systems. The main finding is that carbon domains start nucleating from the amorphous silicon matrix as the temperature goes up during thermal processing.

Mira: And these carbon domains progressively form distinct graphene-like sheets while keeping the overall ceramic network structure intact. They found that this process reproduces the experimental atomic pair distribution functions with really good fidelity, which validates their approach.

Lev: That’s significant because it means they can use this potential to study how those carbon sheets organize themselves during high-temperature annealing cycles, which is something traditional methods struggle with over long timescales.

Kai: But let's look at the training part. They built this interatomic potential using a message-passing neural network architecture called MACE, which has two layers for a good balance between speed and descriptive power across their training database of over nine thousand four hundred forty-three configurations.

Mira: That database itself is pretty diverse, containing seven different classes of configurations, including bulk models, high temperature states, surfaces for the carbon domain boundaries they are interested in studying, and even crystal structure prediction models to sample rare bonding environments.

Lev: A large database helps with generalization on real hardware because it gives the potential a lot of examples to learn from before you throw it into a complex simulation run.

Title and authors: Kai: They showed that this machine learning interatomic potential achieves remarkable accuracy on an unseen test database, with root mean square errors of twelve point four meV per atom for energies and one hundred forty-nine meV per Angstrom for forces.

Mira: Those error numbers suggest the model is capturing the physics quite well, especially when looking at the best two models they identified, SiCNH8000-CL and particularly SiCNH8000-CNB two thousand two hundred <ref:2605.20358#pg1>.

Lev: So what does that mean for us in terms of running simulations on real hardware? If we have an error of about one hundred fifty meV per Angstrom, that’s a pretty tight bound for what we can expect from the simulation results during long runs.

Kai: The analysis shows these best two models are consistent with the presence of two separate phases: a carbon phase containing sp2-like carbon atoms and the silicon carbonitride network. The way their models converge toward experimentally observed structures, no matter where they start, shows how thermodynamically robust this potential is for these complex ceramics.

Mira: Structurally, they looked at the nitrogen local environments and found that it's predominantly three fold coordinated and essentially bonded to silicon atoms as expected in the SiCN network. The carbon structuring goes along with the disappearance of carbon dimers and small chains, leading to sheets containing mainly aromatic rings.

Lev: Tracking those specific structural changes in the carbon motifs is exactly what we need when we want to understand how those domains actually grow during a thermal cycle.

Kai: The phase separation mechanism they describe is that the nucleation and growth of these carbon sheets are mediated by defective carbon rings, which act as intermediate structures. These rings facilitate the formation of stable six member aromatic cycles by capturing extra carbon atoms or ejecting excess ones to optimize local bonding configurations.

Mira: And extending the dynamics and increasing the temperature on their SiCNH8000-CNB two thousand two hundred model allowed them to continue that growth and structuring of free carbon domains that started in the SiCNH8001-CNB model <ref:2605.20358#pg1>.

Title and authors: Lev: So they’re showing a pathway for simulating the long thermal annealing needed to see this biphasic system form, which requires a temperature of two thousand two hundred Kelvin <ref:2605.20358#pg1>. That's a high temperature to handle on many current simulators.

Kai: The PDC global composition can then be decomposed into forty percent free carbon and sixty percent carbon within the SiCN matrix, with a chemical composition of Si32C15N24 without accounting for hydrogen. This chemically consistent phase separation supports the ability of extended machine learning molecular dynamics to model these complex SiCN systems at high temperatures.

Mira: The main implication here is that this methodology can be applied to carbon rich PDC systems to study the formation of extended turbostratic carbon structures, which is something we haven't seen explored with this level of detail before.

Lev: If we can model those extended turbostratic structures, it gives us a much better picture of how the electrical conductivity is boosted by that free carbon phase you mentioned earlier.

Kai: To wrap up on the paper "Modeling phase separation in polymer-derived silicon carbonitride ceramics through extended machine learning molecular dynamics," the convergence of their models toward experimentally observed structures confirms the thermodynamic robustness of their potential for capturing intrinsic structural features and carbon domain growth mechanisms.

Mira: They show that by tuning an initialization and running a long enough simulation time at a temperature deliberately higher than experimental pyrolysis temperature, you get the clear phase separation characterized by dispersed carbon sheets embedded in an SiCN amorphous phase.

Lev: And they emphasize that this biphasic system requires that long thermal annealing at two thousand two hundred Kelvin to actually observe it occurring <ref:2605.20358#pg1>.

Kai: Beyond this specific composition, the methodology they developed can be applied to other carbon rich PDC systems to study the formation of extended turbostratic carbon structures.

Mira: The paper is really about showing how machine learning interatomic potentials can allow us to simulate atomic-scale evolution during thermal treatment in materials like PDCs with a level of detail that was previously inaccessible.

Lev: From an error perspective, they’ve shown that the best model spontaneously evolves toward this carbon phase separation, which is a strong indicator for the utility of these MLIPs in real hardware applications.

The paper's summary: Kai: So basically, this paper uses extended machine learning molecular dynamics to build a potential for silicon carbonitride systems that is good enough to simulate big systems over long times, which is a big deal for us in cryogenics and materials science.

Mira: Right, it takes these complex polymer-derived ceramics, the PDCs, and looks at how they separate into different phases when you heat them up, specifically focusing on that amorphous silicon network versus the free carbon phase.

Kai: The main thing they found is that carbon doesn't just stay as a random mess; it nucleates and grows into these distinct graphene-like sheets right out of the silicon matrix during thermal processing.

Mira: And this happens while still keeping the overall ceramic structure mostly intact, which is important because you want to control that final material microstructure.

Kai: They validated this whole approach by showing their model reproduces experimental pair distribution functions with pretty high accuracy, meaning it’s actually capturing the atomic-scale physics right.

Mira: So for someone just listening, what this means is that we can start using these machine learning potentials to simulate the long-term evolution of these materials under real thermal conditions without needing impossibly slow traditional simulations.

Kai: The training database itself was huge, over nine thousand configurations, covering everything from bulk amorphous material to surfaces where the free carbon domains are starting to form.

Mira: And they used a message passing neural network architecture called MACE for the training because it balances computational speed with how much physical description you get out of it.

Kai: The best models they found, like SiCNH8000-CNB two thousand two hundred consistently show this two-phase system—the carbon phase and the SiCN network—even when they start from different initial conditions.

Mira: Structurally, they looked at the nitrogen environments and confirmed that it’s mostly three fold coordinated to silicon, which is what you expect in a SiCN ceramic network.

Kai: The way the free carbon organizes is by getting rid of those small carbon dimers and chains until you get these larger sheets with mostly aromatic rings.

Mira: And they found the mechanism for this growth involves defective carbon rings acting as intermediates that help form stable six member aromatic cycles by either grabbing extra carbons or spitting out excess ones to optimize how they bond locally.

Kai: They ran simulations on one model, SiCNH8000-CNB two thousand two hundred and by extending the dynamics and pushing the temperature up to two thousand two hundred Kelvin, they saw this free carbon domain actually grow.

Mira: That growth led them to decompose the global composition of the PDC into forty percent of that free carbon and sixty percent within the SiCN matrix, which is chemically consistent.

Kai: The crucial part is that they show this biphasic system needs a long thermal annealing at two thousand two hundred Kelvin to actually happen, not just instantly.

Mira: So if you only listen to this, the big picture is that machine learning tools are becoming powerful enough to model these complex phase separations and their underlying mechanisms in materials like silicon carbonitride for the first time.

The paper's improvements: Tom: So we're talking about how these authors suggest ways to make their machine learning potential even better for studying these materials.

Kai: They suggest improving the scaling factors, specifically the r scaling factor and the amplitude scaling factor, which they think will give us a tighter match with experimental data in the medium range order.

Mira: That makes sense because when you're trying to model phase separation, getting that medium range order right is key to understanding how those domains actually interact.

Kai: They also highlight tracking the evolution of carbon motifs in polymer derived silicon carbonitride ceramics during thermal annealing cycles, showing how those atoms really organize into dimers and chains and sheets.

Mira: That’s important because we don't just want a final structure; we want to see the dynamic process—how it goes from amorphous stuff to organized domains.

Kai: They also mention that the best model evolves toward this clear phase separation on its own if you start it right and run it for long enough at a temperature deliberately higher than what was used experimentally.

Mira: That points back to the thermal annealing requirement, which they say is really necessary to observe that biphasic system forming in the first place.

Kai: So these suggestions focus on getting better accuracy when you're running simulations, specifically focusing on how the potential captures those intermediate structural changes in carbon.

Mira: It means for us, if we are using these ML potentials for phase separation studies, we need to be paying attention to tuning those scaling factors and keeping our simulation times long enough to see the actual reorganization of atoms.

Conclusion: Kai: So we're wrapping up this look at "Modeling phase separation in polymer-derived silicon carbonitride ceramics through extended machine learning molecular dynamics."

Mira: Basically, they showed that their machine learning potential is robust enough to capture the atomic-scale evolution of carbon domains forming in SiCN materials during high temperature treatment.

Kai: And they proved that this process—the nucleation and growth of those sheets—is driven by specific intermediate structures in the carbon rings, which is a mechanism we can now model.

Mira: It really shows that these machine learning methods aren't just good for static images; they can actually map out the dynamics of how materials transform under thermal stress.

Lev: From an engineering standpoint, if this potential works at this accuracy over such long timescales, it means we could finally simulate the actual ceramic processing cycles needed to get those desired properties without having to run simulations that take forever on real hardware.

Kai: Exactly, and their findings on the Si32C15N24 composition decomposing into free carbon and SiCN matrix gives us a concrete chemical picture of what's happening at two thousand two hundred Kelvin.

Mira: That decomposition is powerful because it links the microscopic structure directly to the macroscopic properties we care about, like electrical conductivity.

Lev: I think that ability to model phase separation at such high temperatures, where you need a long annealing time, is what makes this methodology really useful for predicting real-world material behavior under harsh conditions.

Kai: The paper sets a strong foundation for applying this technique to other carbon rich PDC systems, not just this one specific composition they focused on.

Mira: It suggests that we can use these ML potentials as a tool to probe the formation of extended turbostratic carbon structures in those materials too.

Lev: If we can model those extended structures, it gives us a much better picture of how that free carbon phase actually contributes to the material's electrical properties, which is what we need for real-world device design.

Kai: So this paper really confirms that the convergence of their models toward experimental structures validates the thermodynamic robustness of their potential for these complex ceramics.

Mira: It’s a solid piece showing how ML molecular dynamics can uncover these intricate phase separation mechanisms in materials that are otherwise incredibly difficult to study atom by atom.

Lev: It’s a good validation for using AI-driven potentials when you need to study slow, thermally driven structural changes over long simulation times.

More episodes

← Home