A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics".
Title and authors: Tom: So we’ve seen how they built this framework, focusing on learning local rules from limited data, and now we’re looking at what the paper actually says about the title and authors of "A Physics-Regulated Neural Framework for Learning three dee Grain Growth Dynamics" <ref:2607.04680#pg1,A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics>.
Jane: The title itself is pretty descriptive: it tells us exactly what this thing is—it’s a neural framework that's physics-regulated—which suggests they aren't just doing some black box learning; they are incorporating known physical laws into the learning process.
Lu: And the authors, Zhihui Tiana, Kang Yanga, Michael Tonksb, Amanda R. Krausec and Joel B. Harleya from the University of Florida and Carnegie Mellon University, bring a solid mix of expertise in materials science and machine learning representation learning.
Meng: It’s good to see that mix because they clearly understand the physics side—grain growth is governed by boundary energy reduction—and they aren't just throwing a deep learning model at the problem hoping it works.
Tom: And what this means for us, in simple terms, is that we’re moving away from needing huge amounts of perfectly labeled data to simulate complex physical processes like grain growth.
Jane: Instead of needing millions of simulation runs to learn the whole behavior from scratch, we can use a physics-informed approach to create a model that learns the fundamental local interactions directly.
Lu: So, they are showing that you don't need a global representation of the microstructure at every time step to predict how it evolves; locality is key.
Meng: That sounds like something we’ve been struggling with when we try to build models for things that are inherently spatially dependent and huge.
Tom: Precisely, and the paper suggests that by focusing on local interactions, you can achieve accurate prediction without explicitly encoding the entire global system into a neural network's latent space.
Jane: This moves us closer to something where we can predict microstructure evolution efficiently because we only need to capture how one grain boundary interacts with its immediate neighbors.
Lu: The authors are aiming for a surrogate model that captures the physics, and they are trying to ensure that this surrogate model is also computationally scalable for large domains.
Meng: If it's scalable, then it’s actually useful in production because we can deploy it on hardware that can handle really big simulations without massive memory overhead.
Tom: So the paper is about using physics principles to guide a neural network to learn local evolution rules for three dee grain growth dynamics <ref:2607.04680#pg1>.
Jane: And they are showing that this learning process results in a model that is both physically consistent and efficient at scaling up its prediction capabilities.
The paper's summary: Tom: Now let's get into the actual summary of "A Physics-Regulated Neural Framework for Learning three dee Grain Growth Dynamics" <ref:2607.04680#pg1,A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics>. They outline how they set up the experiment, and what they found about their results.
Jane: They start by noting that grain growth follows well-established statistical scaling laws related to the reduction in grain boundary energy, which is why they are trying to build a model that respects those known physical invariants.
Lu: The model is trained using only two consecutive time steps of data, yet it successfully reproduces the linear coarsening law and preserves topological statistics over extended time scales.
Meng: That’s impressive because usually, you need much more than two steps to get any meaningful kinetic information out of a simulation.
Tom: They show that even with this minimal training data, the model accurately reproduces the linear coarsening law and keeps the topological statistics stable over longer prediction horizons.
Jane: They are specifically showing that this model maintains those key statistical features, which proves it’s learning something robust beyond just fitting a short sequence of data points.
Lu: Furthermore, they found that three dee-PRIMME captures local curvature-driven grain boundary migration rules that are transferable across both spatial scales and prediction horizons <ref:2607.04680#pg1>.
Meng: So the local rules they learned aren't just specific to the small simulation box where they trained; they apply generally.
Tom: And one of the key findings is that three dee-PRIMME can reproduce key kinetic, topological, and statistical features of three dimensional grain growth while extrapolating to larger simulation domains <ref:2607.04680#pg1>.
Jane: They also found something interesting about the inclination-dependent training data; this shows the model isn't just memorizing trajectories but actually learning underlying physical principles.
Lu: The results are strong because they demonstrate that three dee-PRIMME can achieve stable grain-growth evolution without needing that extra term, which is referring to the explicit regularization term in the loss function <ref:2607.04680#pg1>.
Meng: That’s a nice win for me; it means the model didn't need some extra hack just to make it work reliably.
Tom: So, basically, they found a way to learn physics-consistent dynamics using extremely limited training data and see results that look really good when extrapolated to much larger systems.
Jane: This framework is designed to be a scalable and data-efficient surrogate for large scale three dee microstructure evolution simulations <ref:2607.04680#pg1>.
The paper's improvements: Tom: Now let's talk about the specific improvements the authors suggest for this work in "A Physics-Regulated Neural Framework for Learning three dee Grain Growth Dynamics" <ref:2607.04680#pg1,A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics>. They aren't just stopping there.
Jane: One improvement they suggest is developing an adaptive window selection strategy, where the size of the observation and action windows changes based on evolving microstructure statistics like average grain size and distribution.
Lu: That would be a smart move because it ensures that the interface-site representation actually captures the most relevant local context without introducing redundant information when you're looking at a very different structure.
Meng: I could see that helping with accuracy, but what about the computational cost of dynamically adjusting those window sizes during a simulation rollout? That sounds like adding complexity to what we just learned.
Tom: That’s a fair point, but they also suggest incorporating an explicit misorientation representation into the framework to make it fully anisotropic, so it accounts for both boundary inclination and other misorientation effects.
Jane: That would be a big step toward making the model applicable to experimental data because you could use three dee-PRIMME as a surrogate for anisotropic grain growth <ref:2607.04680#pg1>.
Lu: And from an engineering standpoint, they also suggest leveraging the local update rule structure, which computes updates independently within local windows to enable highly parallel computation and significantly reduce memory requirements for large scale three dee simulations <ref:2607.04680#pg1>.
Meng: That sounds like something I could actually deploy on modern high performance computing architectures because that locality means updates don't have to be perfectly synchronized across the whole system.
Tom: So, the authors are pushing for improvements that focus on making the model more accurate through better representation and more scalable through parallel computation.
Jane: It seems they are really focused on ensuring this framework is not just a proof of concept but a robust tool for real-world use in simulating materials science.
Conclusion: Tom: So, to wrap up, the main point of "A Physics-Regulated Neural Framework for Learning three dee Grain Growth Dynamics" is that they created a model that can learn local evolution rules from very little data and still maintain physical consistency <ref:2607.04680#pg1,A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics>.
Jane: They showed it can predict key kinetic features like the linear coarsening law and topological statistics reliably, even when extrapolated to much larger simulation domains.
Lu: The paper confirms that this framework is a scalable and data-efficient surrogate for large scale three dee microstructure evolution simulations <ref:2607.04680#pg1>.
Meng: For me, the implication is that we can do high fidelity simulations on large scales without needing astronomical amounts of training data upfront.
Lalam: From my perspective, this means we can build models that capture transferable local update rules rather than just memorizing specific evolution trajectories.
Tom: It’s about learning how things work fundamentally at a lower level, which is what makes these models useful for predicting the future behavior of materials.
Jane: It’s a really solid result that shows physics-informed machine learning can deliver scalable surrogate models for complex problems in three dee microstructure evolution <ref:2607.04680#pg1>.
Zhihui Tiana, Kang Yanga, Michael Tonksb, Amanda R. Krausec, Joel B. Harleya
Department of Electrical and Computer Engineering, University of Florida · Department of Materials Science Engineering, University of Florida · Department of Materials Science Engineering, Carnegie Mellon University
cs.LG, physics.comp-ph
Submitted: 2026-07-06
Updated: 2026-07-06
Journal ref: Materials & Design 270 (2026) 116899
DOI: 10.1016/j.matdes.2026.116899
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The gist: Grain growth is governed by the reduction in grain boundary energy and exhibits well-established statistical scaling laws <ref:2607.04680#pg2> The model is trained using only two consecutive time…
Key concepts
- Local Spatial Windows
- The framework operates on fixed-size local neighborhoods rather than the whole microstructure. This allows it to focus only on the immediate surroundings of a site, making the model inherently scalable to very large 3D domains by using only local information for predictions.
- Interface-Site Representation
- For every site, the framework maps its local neighborhood into a representation based on how many neighboring sites have different grain IDs. This metric encodes the immediate grain-boundary environment, which serves as the primary input for learning how that boundary evolves.
- Hamiltonian Formulation
- The local neighborhood is mapped to an interface-site representation using a Hamiltonian formulation. This mathematical structure helps define the relationship between neighboring sites and their grain boundaries, providing a consistent way to encode the local structural context for the neural network.
Terminology
Summary
Grain growth is governed by the reduction in grain boundary energy and exhibits well-established statistical scaling laws <ref:2607. p. 004680#pg2> The model is trained using only two consecutive time steps yet accurately reproduces the linear coarsening law and preserves topological statistics over extended time scales <ref:2607. p. 004680#pg4> Despite being trained on a 1003 grid points with 512 grains, the learned evolution operator is applied to domains up to 10243 grid points with 550000 grains without retraining, maintaining consistent kinetics and grain topology across orders-of-magnitude increases in system size <ref:2607. p. 004680#pg2> These results demonstrate that 3D-PRIMME learns a scale-independent and temporally stable local evolution rule, enabling efficient and robust large-scale surrogate prediction of 3D microstructure evolution <ref:2607. p. 004680#pg2>.
Framework Design
The framework is designed to learn the local evolution rules governing 3D microstructure dynamics from data <ref:2607.04680#pg3> Rather than relying on a global representation of the microstructure, the framework operates on fixed-size local spatial windows, which makes it naturally scalable to large 3D domains <ref:2607. p. 104680#pg3>. For each target site, the local neighborhood is mapped to an interfacesite representation following the Hamiltonian formulation <ref:2607. p. 104680#pg4> This structural metric counts neighboring sites with different grain IDs within the observation window, thereby encoding the local grain-boundary environment that serves as the model input <ref:2607. p. 104680#pg4>.
Training Data Construction
The training and evaluation datasets are generated using the mode-filter (MF) grain growth model <ref:2607.04680#pg4> The MF model is adopted for three main reasons, including introducing fluctuations in the evolving microstructures to provide a useful proxy for experimental noise <ref:2607. p. 104680#pg4>. Training data are constructed by sampling M sequences and 2 consecutive time steps from the MF simulations to construct different training sets <ref:2607. p. 204680#pg4>. The resulting datasets have the shape (M, 2, 100, 100, 100), where M is a hyperparameter that controls the amount of training data
Performance and Generalization
The model demonstrates strong capability in both spatial and temporal extrapolation The evolution of the squared average grain radius, ⟨r⟩2 is nearly identical for the three domains, indicating consistent coarsening kinetics across different system sizes The model captures local curvature-driven grain boundary migration rules that are transferable across both spatial scales and prediction horizons
Key Findings on Dynamics
The results demonstrate that 3D-PRIMME can reproduce key kinetic, topological, and statistical features of threedimensional grain growth while extrapolating to larger simulation domains The model achieves stable grain-growth evolution without this additional term [referring to the explicit regularization term in the loss function]. Furthermore, 3D-PRIMME is able to reproduce inclination-dependent grain growth behavior from anisotropic training data. p. 5
Conclusion
The present framework provides a scalable and data-efficient surrogate for large-scale 3D microstructure evolution The model captures local curvature-driven grain boundary migration rules that are transferable across both spatial scales and prediction horizons. The inclination-dependent grain growth results provide further evidence that 3D-PRIMME learns transferable local update rules rather than simply memorizing evolution trajectories. p. 7 p. 7
Data Availability
The data and code supporting the findings of this study will be made available in a public repository upon publication The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgements
The authors would like to acknowledge financial support by the U.S. Department of Energy, Office of Science, Basic Energy Sciences under Award DE-SC0020384 The references listed are [1] F. p. 0-Y. Wu, The potts model, Reviews of modern physics 54 (1982) 235. p. 0
References
[1] F.-Y. Wu, The potts model, Reviews of modern physics 54 (1982) 235
[2] L.-Q. Chen, Phase-field models for microstructure evolution, Annual review of materials research 32 (2002) 113–140
[3] Y. He, H. Ding, L. Liu, K. Shin, Computer simulation of 2d grain growth using a cellular automata model based on the lowest energy principle, Materials Science and Engineering: A 429 (2006) 236–246
[4] E. A. Lazar, J. K. Mason, R. D. MacPherson, D. J. Srolovitz, A more accurate three-dimensional grain growth algorithm, Acta Materialia 59 (2011) 6837–6847
[5] J. Fausty, B. Murgas, S. Florez, N. Bozzolo, M. Bernacki, A new analytical test case for anisotropic grain growth problems, Applied Mathematical Modelling 93 (2021) 28–52
[6] M. A. Miodownik, A review of microstructural computer models used to simulate grain growth and recrystallisation in aluminium alloys, Journal of Light Metals 2 (2002) 125–135
[7] W. Yan, J. Melville, V. p. 1 Yadav, K. Everett, L. Yang, M. S. Kesler, A. R. Krause, M. R Tonks, J B Harley, A novel physics-regularized interpretable machine learning model for grain growth
[8] S Fan, A L Hittt, M Tang, B Sadigh, F Zhou, Accelerate microstructure evolution simulation using graph neural networks with adaptive spatiotemporal resolution, Machine Learning: Science and Technology 5 (2024) 025027
[9] K Yang, Y Cao, Y Zhang, S Fan, M Tang, D Abergt B Sadigh, F Zhou, Self-supervised learning and prediction of microstructure evolution with convolutional recurrent neural networks Patterns 2 (2021)
[10] I Peivaste, A Makradi, S Belouettar, Teaching artificial intelligence to perform rapid, resolution-invariant grain growth modeling via fourier neural operator Computer Methods in Applied Mechanics and Engineering 440 (2025) 117945
[11] P Tep, M Bernacki, High-fidelity grain growth modeling: Leveraging deep learning for fast computations arXiv preprint arXiv:2505.05354 (2025)
[12] Z Tian, E Suwandi, T Oppelstrup, V V Bulatov, J B Harley, F Zhou Scaling kinetic monte carlo simulations of grain growth with combined convolutional and graph neural networks arXiv preprint arXiv:2511.17848 (2025)
[13] W Ludwig, A King, P Reischig, M Herbig, E M Lauridsen, S Schmidt, H Proudhon, S Forest P Cloetens S Du Roscoat et al., New opportunities for 3d materials science of polycrystalline materials at the micrometre lengthscale by combined use of x-ray diffraction and x-ray imaging Materials Science and Engineering: A 524 (2009) 69–76
[14] E A Holm, P M Duxbury, Three-dimensional materials science Scripta materialia 54 (2006) 1035–1040
[15] V Yamakov, D Moldovan, K Rastogi, D Wolf, Relation between grain growth and grain-boundary diffusion in a pure material by molecular dynamics simulations Acta materialia 54 (2006) 4053–4061
[16] Z Xu, Y.-F Shen, S K Naghibzadeh, X Peng, V Muralikrishnan, S Maddali, D Menasche A R Krause K Dayal R M Suter et al., Grain boundary migration in polycrystalline α-fe Acta Materialia 264 (2024) 119541
[17] J Mason, J Lind, S Li, B Reed, M Kumar, Kinetics and anisotropy of the monte carlo model of grain growth Acta Materialia 82 (2015) 155–166
[18] M Anderson, D Srolovitz, G Grest, P Sahni, Computer simulation of grain growth—i. kinetics Acta metallurgica 32 (1984) 783–791
[19] J Melville, V Yadav, L Yang, A R Krause, M R Tonks, J B Harley A new efficient grain growth model using a random gaussian-sampled mode filter Materials & Design 237 (2024) 112604
[20] F Wakai, N Enomoto, H Ogawa, Three-dimensional microstructural evolution in ideal grain growth—general statistics Acta Materialia 48 (2000) 1297–1311
[21] R D Kamachali, I Steinbach, 3-d phase-field simulation of grain growth: Topological analysis versus mean-field approximations Acta Materialia 60 (2012) 2719–2728
[23] E Miyoshi, T Takaki, M Ohno, Y Shibuta, S Sakane, T Shimokawabe, T Aoki Ultra-large-scale phase-field simulation study of ideal grain growth NPJ Computational Materials 3 (2017) 25
[24] J Hestness, S Narang, N Ardalani, G Diamos, H Jun, H Kianinejad, M M A Patwary Y Yang Y Zhou Deep learning scaling is predictable empirically arXiv preprint arXiv:1712.00409 (2017)
[25] S Mindermann, J M Brauner, M T Razzak, M Sharma, A Kirsch, W Xu B Höltgen A N Gomez A Morisot S Farquhar et al., Prioritized training on points that are learnable, worth learning, and not yet learnt in: International Conference on Machine Learning PMLR 2022 pp. p. 6 15630–15649
[26] L Yang, V Yadav, J Melville, J B Harley A triple junction energy study using an inclination-dependent anisotropic monte carlo potts grain growth model Materials & Design 239 (2024) 112763
[27] L Yang, F Hilty, V Muralikrishnan, K Silva-Reyes, J B Harley A R Krause M R Tonks Calculating the grain boundary inclination of voxelated grain structures using a smoothing algorithm Scripta Materialia 218 (2022) 114796
[28] E Hüllermeier, W Waegeman, Aleatoric and epistemic uncertainty in machine learning: An introduction to concepts and methods Machine learning 110 (2021) 457–506
[29] M McCabe, P Harrington, S Subramanian, J Brown, Towards stability of autoregressive neural operators arXiv preprint arXiv:2306.10619 (2023)
[30] B List, L.-W Chen, K Bali, N Thuerey, Differentiability in unrolled training of neural physics simulators on transient dynamics Computer Methods in Applied Mechanics and Engineering 433 (2025) 117441
[31] D Zhuang, X Zhang, S Song, S Hooker, Randomness in neural network training: Characterizing the impact of tooling Proceedings of Machine Learning and Systems 4 (2022) 316–336
[32] J Melville, V Yadav, L Yang, A R Krause, M R Tonks J B Harley Anisotropic physics-regularized interpretable machine learning of microstructure evolution Computational Materials Science 238 (2024) 112941
[S1] Quantitative comparison of observation and action window sizes. The relative error in ⟨r⟩2 and the MAE in the average number of grain faces are computed relative to the MF reference over the 100-step rollout
[S2] Statistical comparison between MF and 3D-PRIMME predictions with uncertainty quantification when only the data-shuffling seed is varied across runs, while the random initialization is fixed. (a) Mean squared grain size, ⟨r⟩2, as a function of step
[S3] Comparison of stereographic projections of inclination distributions for the MF reference, top row, and 3DPRIMME prediction, bottom row, at simulation steps 0, 1, 10, and 50
Pruning redundant reasoning tokens
The observation window size has a strong influence on the growth rate An observation window of N o = 9 results in the most accurate reproduction of the MF growth rate The relationship between the average number of grain faces with number of time steps also remains fairly consistent with the MF results This suggests that the observation window has a stronger influence on the learned dynamics because it directly determines the local neighborhood used to construct the interface-site representation.
How the training signal is built
The training target is constructed by checking whether site i adopts the current grain label of a candidate site j in the action window after one evolution step For each candidate site j ∈ Δa(i), a binary label y(t)i j is defined as 1 if s(t+1)i = s(t)j, and 0 otherwise. A squared-error loss is used to link the interface-site inputs to the predicted action likelihoods
Where the gains come from
The model can learn physically consistent grain growth dynamics using extremely limited training data The model captures local curvature-driven grain boundary migration rules that are intrinsically scale-independent. This locality enables efficient deployment on large 3D microstructures, supporting simulations at spatial scales far beyond those used during training. p. 1 The inclination-dependent grain growth results provide further evidence that 3D-PRIMME learns transferable local update rules rather than simply memorizing evolution trajectories.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Adaptive window selection strategy: Develop an adaptive window-selection strategy where observation and action window sizes are adjusted according to evolving microstructural statistics, such as average grain size, grain-size distribution, and number of grains. This would improve prediction accuracy by ensuring the interface-site representation captures the relevant local structural context without introducing redundant information.
-
Explicit misorientation representation: Incorporate an explicit misorientation representation into the 3D framework to enable a fully anisotropic model that accounts for both boundary inclination and misorientation effects. This extension will provide a path toward applying PRIMME to experimentally measured 3D microstructures and using it as a surrogate model for anisotropic grain-growth evolution.
-
Scalable parallel computation: Leverage the local update rule structure, which computes updates independently within local windows, to enable highly parallel computation and significantly reduce memory requirements for large scale 3D simulations. This locality allows efficient deployment on modern high performance computing architectures for simulations at spatial scales far beyond those used during training.
Sources
- High-fidelity Grain Growth Modeling: Leveraging Deep Learning for Fast Computations
- Scaling Kinetic Monte-Carlo Simulations of Grain Growth with Combined Convolutional and Graph Neural Networks
- Deep Learning Scaling is Predictable, Empirically
- Towards Stability of Autoregressive Neural Operators
Related papers
- OptiMind: Improving Small Language Models for Optimization Modeling through Formulation Guidance and Multi-Turn Learning
- Staying in the Sweet Spot: Boost Reasoning Evolution via Responsive Rollout Difficulty Manipulation
- Examining the Impact of Data and Training Methods on Emergent and Subliminal Misalignment
- Rethinking Black-Box On-Policy Distillation: A Rubric-based Analysis
- (Mis)generalization of Helpful-Only Fine-tuning
- Provably Learning Diffusion Models under the Manifold Hypothesis: Collapse and Refine