Neural surrogates for crystal growth dynamics with variable supersaturation: explicit vs. implicit conditioning

summary

The gist

Simulations of crystal growth are performed by using Convolutional Recurrent Neural Network surrogate models…

In short

This study compares two neural network models for simulating crystal growth dynamics under varying supersaturation. One model inferred supersaturation implicitly from short sequences, while the other explicitly took it as input. The results show that explicitly conditioning on the known supersaturation value consistently yields higher prediction quality and requires significantly less training data, making explicit conditioning the superior approach.

Key concepts

Allen-Cahn dynamics
This is a mathematical model used to simulate how materials grow or evolve over time, similar to how crystals form. It describes the movement of a field ($ϴ$ map) that changes based on energy minimization, which dictates the shape and structure of the growing crystal surface.
Convolutional Recurrent Neural Network (CRNN)
This is a type of deep learning model designed to analyze sequences of data, such as time-series images or simulation frames. The convolutional part extracts spatial features from each frame, while the recurrent part processes the sequence order to understand how the crystal evolves over time.
Supersaturation ($ΔƵ$)
This is a critical physical parameter in crystal growth that measures how far the system is from equilibrium. It dictates whether crystals grow, shrink, or remain stable. The study investigates how different supersaturation levels affect the accuracy of predicting crystal shapes.
Explicit vs. Implicit Conditioning
Explicit conditioning means directly feeding a known parameter (like the exact supersaturation value) into the AI model as an input. Implicit inference means letting the model try to figure out that same parameter by looking at only a short sequence of evolution frames. The paper proves explicit input is better for reliable results.

Terminology used across episodes

This episode discusses

The paper

Neural surrogates for crystal growth dynamics with variable supersaturation: explicit vs. implicit conditioning · Read on arXiv

University of Milano-Bicocca · University of Genova

Simulations of crystal growth are performed by using Convolutional Recurrent Neural Network surrogate models, trained on a dataset of time sequences computed by numerical integration of Allen-Cahn dynamics including faceting via kinetic anisotropy. Two network architectures are developed to take into account the effects of a variable supersaturation value. The first infers it implicitly by processing an input mini-sequence of a few evolution frames and then returns a consistent continuation of the evolution. The second takes the supersaturation parameter as an explicit input along with a single initial frame and predicts the entire sequence. The two models are systematically tested to establish strengths and weaknesses, comparing the prediction performance for models trained on datasets of different size and, in the first architecture, different lengths of input mini-sequence. The analysis of point-wise and mean absolute errors shows how the explicit parameter conditioning guarantees the best results, reproducing with high-fidelity the ground-truth profiles. Comparable results are achievable by the mini-sequence approach only when using larger training datasets. The trained models show strong conditioning by the supersaturation parameter, consistently reproducing its overall impact on growth rates as well as its local effect on the faceted morphology. Moreover, they are perfectly scalable even on 256 times larger domains and can be successfully extended to more than 10 times longer sequences with limited error accumulation. The analysis highlights the potential and limits of these approaches in view of their general exploitation for crystal growth simulations.

DOI: 10.1088/2632-2153/aeb226

Transcript

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

Kai: We're looking at the paper: "Neural surrogates for crystal growth dynamics with variable supersaturation".

Mira: Here's the gist of the paper. Simulations of crystal growth are performed by using Convolutional Recurrent Neural Network surrogate models, trained on a dataset of time sequences computed by numerical integration of Allen-Cahn dynamics including faceting via kinetic anisotropy The study finds that explicitly conditioning on the supersaturation value reliably increases the quality of predictions and yields more data-efficient NN models Model Architectures Compared Two network architectures were developed to account for variable supersaturation, one inferring it implicitly by processing a mini-sequence of evolution frames, and the second taking the supersaturation parameter as an explicit input along with a single initial frame The first model is called NNseq, which acts as a sequence-completion task by requiring multiple timeframes as input for the NN model The second model, NNpar, receives the actual value of the supersaturation parameter explicitly by converting it into a uniform spatial tensor and concatenating it to the φ map Training Methodology A dataset of 7500 PF simulation sequences was constructed for training both NN models, where each case consisted of 200 time-frames reporting the φ field as evolving from a random distribution of elliptical seeds toward complete coverage The loss function L is defined as the common mean squared error (MSE) between the PF ground-truth and NNpredicted profiles, averaged on all time-frames t of each sequence and on all Nb samples in the training batch Curriculum learning was implemented, starting from predicting only the last frame of a sequence and gradually reducing the number of input frames for NNseq Performance Gains The study found that explicitly conditioning on the supersaturation value reliably increases prediction quality and yields more data-efficient models, requiring approximately 15 times larger datasets for NNseq exploiting small sequences compared to explicitly conditioned ones The best performing model, NN5k∆µ, is capable of producing quantitative predictions with local errors limited to just a few pixels (more than 90% of the tested cases are predicted with less than 1.8% of wrong pixels) NNpar models performed significantly better than NNseq ones, showing that explicit conditioning strongly simplifies the task with respect to inferring it implicitly Generalization and Extrapolation The trained models were successfully applied to larger computational domains, with no loss of accuracy, demonstrating potential for application on any domain size The NN can reliably predict the evolution of profiles starting from initial coverage as low as 0.03, and its performance remains consistent across all sizes when compared to the training conditions Extrapolation toward higher supersaturation beyond 0.8 is possible, limiting the accessible range to ∆µ ≤ 1.0 The NN5k∆µ model outperforms NN7.5k∆µ for any ∆μ within the training range, with maximum MAE errors remaining well below 0.04 Explicit conditioning by supplying the critical parameter as input ensures the best predictions, yielding the lowest mean absolute error in the cases analysed The supremacy of NNpar was expected as the explicit conditioning by supplying the correct ∆µ value as input strongly simplifies the task with respect to inferring it implicitly from a short mini-sequence The similarity between NN0.5k∆µ and NN7.5k∆µ indicates that the performance gap of NNseq can be closed only at the price of a more demanding training on significantly larger datasets The trained models were also successfully applied to larger computational domains, with no loss of accuracy The numerical solution of the present Allen-Cahn does not pose severe computational bottlenecks Data Availability The datasets used to train and test the model are openly available in Materials Cloud Archive at https://doi.org/10.24435/materialscloud:yv-sy The code used to train the NN model is freely available on GitHub at https://github.com/dlanzo/CRANE The datasets used to train and test the model are openly available in Materials Cloud Archive at https://doi.org/10.24435/materialscloud:yv-sy The code used to train the NN model is freely available on GitHub at https://github.com/dlanzo/CRANE The datasets used to train and test the model are openly available in Materials Cloud Archive at https://doi.org/10.24435/materialscloud:yv-sy The code used to train the NN model is freely available on GitHub at https://github.

Title and authors: Kai: So Mira, this paper, "Neural surrogates for crystal growth dynamics with variable supersaturation: explicit vs. implicit conditioning," it looks at using these convolutional recurrent neural network models to speed up crystal growth simulations by replacing the slow numerical integrations of Allen-Cahn dynamics.

Mira: Right, and the core thing they are testing is whether it's better to let the neural network figure out that variable supersaturation on its own, or if we explicitly feed that value into the model.

Lev: From my side, I'm thinking about how much of that speedup matters when you actually try to run this on real hardware; if we rely on implicit inference, it might just mean we're baking in some hidden assumptions about the growth conditions.

Kai: Exactly, and the paper sets up two models: one called NNseq that infers the supersaturation implicitly by looking at a short sequence of evolution frames, and another one called NNpar that takes the actual supersaturation value as an explicit input along with just one starting frame.

Mira: The summary of this paper is really focused on comparing these two approaches using a dataset of seven thousand five hundred simulation sequences where the phi field evolves from random elliptical seeds toward complete coverage.

Lev: So they are looking at how prediction quality changes when you condition the network explicitly versus letting it infer that critical parameter from limited input data.

Kai: The main finding is that explicitly conditioning on the supersaturation value reliably increases prediction quality and actually makes the neural network models more data-efficient in terms of what they need to learn.

Mira: They found that for NNseq, if you want the same level of accuracy, you need about fifteen times larger datasets compared to the explicitly conditioned models.

Lev: That's a big caveat for anyone building this for real hardware; it suggests that relying on implicit inference is really pushing the limits of what we can learn efficiently without massive amounts of training data.

Kai: The best performing model they found, NN5kDeltaMu, can produce quantitative predictions with local errors limited to just a few pixels in more than ninety percent of tested cases, meaning less than one point eight percent wrong pixels overall.

Mira: That level of accuracy across the whole spectrum is what's interesting because it shows this explicit conditioning really simplifies the task compared to letting the network infer it implicitly.

Lev: And when we look at extrapolation, they showed these trained models can be applied to much larger computational domains without losing accuracy, and they can reliably predict evolution starting from very low coverage values as low as zero point zero three.

Kai: They also tested pushing the supersaturation beyond one point zero in some cases, showing it's possible up to a value of one point zero, but beyond that range is limited.

Mira: The paper suggests that for any supersaturation value within the training range, the NN5kDeltaMu model outperforms the NN7 point 5kDeltaMu model with maximum mean absolute error staying well below zero point zero four <ref:2604.21753#pg1>.

Lev: So it seems explicit conditioning really gives us a more stable and efficient convergence of the model, which is what I'm hoping for when we move from theory to actual computation on hardware.

Kai: To wrap up, this paper on "Neural surrogates for crystal growth dynamics with variable supersaturation: explicit vs. implicit conditioning" shows that explicit conditioning provides the best results and the lowest mean absolute error in their analysis of those seven thousand five hundred simulation sequences.

Mira: The main implication is that for predicting crystal growth dynamics, explicitly feeding parameters like supersaturation into a convolutional recurrent neural network model is strongly preferred over letting it implicitly discover them from a mini-sequence of evolution frames.

Lev: From an engineering standpoint, this means we can build models that are more robust and require less data to get reliable predictions for these complex physical processes.

Kai: So, the explicit parameter conditioning guarantees the best results by reproducing the ground-truth profiles with high fidelity across a wide range of conditions.

Mira: And they also showed that while NNseq can be closed at the price of significantly larger datasets, NNpar models performed much better than NNseq ones because explicit conditioning strongly simplifies what’s being asked of the model.

Lev: I just want to say that the similarity between models like NN0 point 5kDeltaMu and NN7 point 5kDeltaMu points to how much effort it would take to close that performance gap in implicit inference, suggesting we need those larger datasets for that route <ref:2604.21753#pg1>.

Kai: It’s clear they are pointing toward using explicit input as the way forward for high-fidelity surrogate modeling in this area of materials science.

Mira: So, the idea is that we don't need the network to guess the physical conditions when it can be given them directly, which leads to better prediction quality and efficiency.

Lev: We’ll keep an eye on how these explicit conditioning methods scale up when we try to map this onto real-time simulation pipelines for things like defect formation or phase transitions.

Kai: That's what we'll be watching as we look at the next set of papers.

The paper's summary: Kai: So, to recap, this paper is looking at using neural networks to speed up crystal growth simulations by comparing two ways to tell the network about how much supersaturation there is—either letting it guess it on its own, or explicitly telling it the number every time.

Mira: Right. It boils down to this trade-off between implicit inference and explicit conditioning when you use these convolutional recurrent neural networks for crystal growth dynamics.

Lev: From my side, I'm seeing that when you make the network infer the supersaturation implicitly from a short sequence of frames, it really forces you to train it on way more data just to get reasonable results.

Kai: That makes sense because when we look at the numbers, for every fifteen times more data you need for the implicit version, the explicitly conditioned model—the NNpar—can handle much bigger problems with a lot less training material.

Mira: Exactly. The explicit conditioning is like giving the network a map before it starts driving; it stabilizes the whole process and makes convergence way more efficient, which is what we want when you’re dealing with these complex physical systems.

Lev: And that efficiency matters because trying to run those implicit models on real hardware would just mean running them for a longer time, increasing the error accumulation across those sequential steps.

Kai: So what they found is that explicitly feeding the supersaturation value in makes the prediction quality much higher and way more reliable, even when dealing with larger simulation domains or starting from very low coverage values.

Mira: They showed that this explicit input doesn't just give better accuracy; it also lets the model handle a wider range of conditions, even pushing beyond one point zero in some tests.

Lev: That’s important because it means this surrogate model isn't just good for one specific growth condition; it’s more versatile.

Kai: And they found that the best performing version, NN5kDeltaMu, can actually give quantitative predictions with errors limited to just a few pixels in most cases.

Mira: So the big picture here is that for this type of physics simulation, explicitly conditioning on the critical physical parameter is strongly preferred over letting the network figure it out automatically.

Lev: It simplifies the task significantly when you look at how much data you need, which is a major win for making these tools practical for real computational error correction setups.

Kai: So what this means practically is that if you're building a simulator, don't just let the network guess the growth environment; feed it the known conditions directly.

Mira: It leads to more stable and efficient convergence of those neural networks, which translates into better prediction quality across a much broader spectrum of physical situations.

Lev: We should focus on architectures that support this explicit input because that's where we can actually see a reliable performance gain when we try to map this onto real hardware constraints.

The paper's improvements: Tom: So, to wrap up, the authors of this paper are suggesting that explicitly conditioning on supersaturation is just better than letting the network discover it implicitly from a short sequence of frames.

Kai: They are pushing for explicit input because they found that it gives you much more stable and efficient convergence when training these neural networks for growth dynamics.

Mira: It’s about preferring explicit information over the automatic correlation discovery capabilities of deep learning methods, which they say leads to better prediction quality overall.

Lev: From a hardware standpoint, that explicit conditioning is what makes it much easier to run on real hardware because it reduces the overhead and stabilizes the training process significantly.

Kai: They quantified this by showing that for models trying to infer supersaturation implicitly, you need about fifteen times more data just to get the same level of accuracy as when you condition it explicitly.

Mira: That’s a big number, and it really highlights how much simpler the task becomes when you supply that parameter directly instead of making the network hunt for it in a sequence.

Lev: It means we can potentially build more robust models that require less massive datasets to get reliable results for these crystal growth simulations.

Kai: And they also found that their best performing model, NN5kDeltaMu, is really capable of giving quantitative predictions across the whole tested range with errors limited to just a few pixels in over ninety percent of cases.

Mira: That level of precision across the spectrum is what shows that this explicit conditioning isn't just theoretical; it actually delivers high-fidelity results for these evolution sequences.

Lev: It’s good to know that when we look at generalization, these models can be extended to larger computational domains without any loss of accuracy, which makes them much more useful for real-world materials science problems.

Kai: They also showed that the network can reliably predict how profiles evolve starting from really low coverage values as low as zero point zero three, regardless of the supersaturation level in those tested cases.

Mira: That reliability across different conditions is what makes this approach so interesting for modeling these complex physical processes in a broader sense.

Lev: So the main improvement they are suggesting is that we should move toward models that take parameters explicitly as input because it dramatically simplifies what the network has to figure out on its own.

Kai: It’s about making sure the input information is there when you need it, rather than hoping a sequence of frames will accidentally reveal the exact supersaturation level needed for the growth.

Conclusion: Kai: So we've been talking about how explicitly conditioning on supersaturation gives us better performance and more data efficiency when using convolutional recurrent neural network surrogates for crystal growth dynamics with variable supersaturation: explicit vs. implicit conditioning, right?

Mira: Exactly. The main point is that feeding the critical parameter directly into the model yields higher fidelity predictions and makes the whole system more stable than trying to let the network infer it on its own from a short sequence of frames.

Lev: From an error correction standpoint, that explicit input cuts down on how much data you actually need to train those models for real-world hardware applications.

Kai: It means we’re moving away from models that might just be memorizing patterns in the sequence and toward ones that are actually learning the underlying physics more directly.

Mira: It's about ensuring the network doesn't have to guess a physical constant when it can be given it clearly, which leads to much better accuracy across a wider range of conditions.

Lev: And for running this on actual hardware, that stability is crucial because we don’t want the model struggling with noisy or unpredictable inputs from the simulation data.

Kai: So, in the end, this paper strongly suggests that explicit conditioning is the way to go when modeling these kinds of complex growth dynamics.

Mira: It shows that for our field, preferring explicit input over implicit inference leads to lower mean absolute errors and more reliable results for simulating crystal evolution.

Lev: I just want to say that it’s a necessary step if we want these tools to actually be usable in any kind of computational experiment down the line.

Kai: So, that wraps up our discussion on how explicit conditioning improves the neural surrogates for crystal growth dynamics with variable supersaturation: explicit vs. implicit conditioning.

Mira: It really reinforces that preference for giving the model what it needs to know directly instead of forcing it to discover things implicitly from limited input data.

Lev: It’s a solid finding because it shows a clear path toward building more efficient and robust simulation tools for materials science applications.

More episodes

← Home