Predicting Grain Growth in Polycrystalline Materials Using Deep Learning Time Series Models
Eliane Younes, Elie Hachem, Marc Bernacki · 2025
Read Predicting Grain Growth in Polycrystalline Materials Using Deep Learning Time Series Models by Eliane Younes, Elie Hachem, Marc Bernacki (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-11630.txt.
Introduction
Eliane Younes, Elie Hachem, and Marc Bernacki compare deep learning models for grain growth forecasts. Grain growth shapes how strong a material is. Full physics simulations capture the growth but cost far too much. Cheaper prediction tools are needed for microstructure design.
Four model families are tested on the same data. They are recurrent networks, long short-term memory, temporal convolutional networks, and transformers. The data come from high-fidelity simulations, reduced to simple statistical summaries. One hundred twenty growth sequences become grain size curves over time. Each model forecasts future curves from a short past, step by step. The comparison shows which models track grain change most faithfully.
Catalog record
| CREATOR | Eliane Younes, Elie Hachem, Marc Bernacki |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 49 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-11630
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Predicting Grain Growth in Polycrystalline Materials Using Deep Learning Time Series Models Eliane Younes, Elie Hachem, Marc Bernacki Abstract Grain Growth strongly influences the mechanical behavior of materials, making its prediction a key objective in microstructural engineering. In this study, several deep learning approaches were evaluated, including recurrent neural networks (RNN), long short-term memory (LSTM), temporal convolutional networks (TCN), and transformers, to forecast grain size distributions during grain growth. Unlike full-field simulations, which are computationally demanding, the present work relies on mean-field statistical descriptors extracted from high-fidelity simulations. A dataset of 120 grain growth sequences was processed into normalized grain size distributions as a function of time. The models were trained to predict future distributions from a short temporal history using a recursive forecasting strategy. Among the tested models, the LSTM network achieved the highest accuracy (above 90%) and the most stable performance, maintaining physically consistent predictions over extended horizons while reducing computation time from about 20 minutes per sequence to only a few seconds, whereas the other architectures tended to diverge when forecasting further in time. These results highlight the potential of low-dimensional descriptors and LSTM-based forecasting for efficient and accurate microstructure prediction, with direct implications for digital twin development and process optimization. Keywords: Grain Growth, Deep Learning, Time Series Forecasting, Microstructural Evolution, Mean-Field Descriptors 1 Introduction Although the grain growth (GG) mechanism has been known for more than one century and actively studied for 70 years [ 62 , 12 , 47 , 67 ] , it remains a very active research topic in terms of in-situ data acquisition [ 10 ] , involved equations, and modelling [ 55 ] . This interest can be directly related to the importance of predicting grain size during annealing of metallic materials, which acts as a first-order parameter in discussing certain mechanical properties such as flow stress or creep resistance [ 75 , 39 ] . The historical background has brought us to the current global picture [ 27 ] : GG in single-phase polycrystals relatively free of point and line defects is driven by the minimisation of grain boundary (GB) areas to decrease the system’s interfacial energy. The velocity \mathbf{v} at which GBs migrate when subjected to a driving pressure P is assumed to be the product of the GB mobility \mu and P\mathbf{n} with \mathbf{n} the outward unitary normal to the GB. Following the chain rules clearly depicted by Herring in [ 27 ] , the capillarity driving pressure linked to an evolving GB interface can be described, at the first order of the small displacement performed, as: P\approx-\underbrace{\left(\gamma\boldsymbol{\mathbb{Id}}+\nabla_{\mathbf{n}}\nabla_{\mathbf{n}}\gamma\right)}_{\boldsymbol{{\mathchoice{\raisebox{0.0pt}{\scalebox{1}[-1]{$\displaystyle\mathbb{L}$}}}{\raisebox{0.0pt}{\scalebox{1}[-1]{$\textstyle\mathbb{L}$}}}{\raisebox{0.0pt}{\scalebox{1}[-1]{$\scriptstyle\mathbb{L}$}}}{\raisebox{0.0pt}{\scalebox{1}[-1]{$\scriptscriptstyle\mathbb{L}$}}}}}\left(\mathbf{n}\right)}:\boldsymbol{\mathbb{K}}, (1) with \boldsymbol{\mathbb{K}}=\nabla\mathbf{n} the curvature tensor and \boldsymbol{{\mathchoice{\raisebox{0.0pt}{\scalebox{1}[-1]{$\displaystyle\mathbb{L}$}}}{\raisebox{0.0pt}{\scalebox{1}[-1]{$\textstyle\mathbb{L}$}}}{\raisebox{0.0pt}{\scalebox{1}[-1]{$\scriptstyle\mathbb{L}$}}}{\raisebox{0.0pt}{\scalebox{1}[-1]{$\scriptscriptstyle\mathbb{L}$}}}}}\left(\mathbf{n}\right) the GB stiffness tensor. Beyond the temperature dependence of the mobility, \mu and \gamma , depend also on the underlying crystallography at GBs which are today depicted in the 5-dimensional parameter space of the misorientation of the abutting crystals and the GB plane (inclination). This leads to significant spatial heterogeneity of GB kinetic properties in polycrystalline materials [ 62 , 36 ] . Moreover, generally, \boldsymbol{{\mathchoice{\raisebox{0.0pt}{\scalebox{1}[-1]{$\displaystyle\mathbb{L}$}}}{\raisebox{0.0pt}{\scalebox{1}[-1]{$\textstyle\mathbb{L}$}}}{\raisebox{0.0pt}{\scalebox{1}[-1]{$\scriptstyle\mathbb{L}$}}}{\raisebox{0.0pt}{\scalebox{1}[-1]{$\scriptscriptstyle\mathbb{L}$}}}}}\left(\mathbf{n}\right) and \boldsymbol{\mathbb{K}} are not diagonalizable in the same basis of orthonormal eigenvectors and the calculation of the tensor double product contracted is not straightforward and must be carefully evaluated in the right basis by also taking into account the dependence of \gamma to crystallographic orientation symmetries [ 1 , 16 ] . Finally, when \gamma is assumed not dependent of the inclination \mathbf{n} (non-anisotopic), i.e. only dependent of the GB disorientation angle (heterogeneous) or homogeneous, Eq. 1 simplifies in the well-known mesoscopic kinetic equation for curvature flow: P\approx-\gamma\left(\kappa_{1}+\kappa_{2}\right)=-\gamma\kappa,\text{ and }\mathbf{v}\approx-\mu\gamma\kappa\mathbf{n}, (2) with \left\{\kappa_{1},\kappa_{2}\right\} the principal curvatures of the curvature tensor \boldsymbol{\mathbb{K}} and \kappa=\kappa_{1}+\kappa_{2} its trace, generally known as the mean curvature of an interface in materials science and, in a mathematically more consistent way, as twice the mean curvature in other communities such as fluid mechanics. In this formulation, the contributions of \mu and \gamma can no longer be clearly separated, and the term reduced mobility has therefore become widely adopted in the literature to denote their product. Thanks to the improvement of GG experiments, this global picture is more and more discussed in the state of the art. Indeed, monitored in-situ in a synchrotron by 3D X-ray diffraction microscopy techniques, which avoid bias inherent to 2D observations, opens the way to precise reverse engineering in order to discuss the \left\|\mathbf{v}\right\|/\kappa ratio [ 13 , 10 , 70 , 20 , 43 , 71 , 48 , 11 , 76 ] , i.e., the apparent reduced mobility of each GB while knowing their characteristics in the 5-dimensional parameter space of the misorientation and inclination. If Eq. 2 is indeed a simplification of lower scale phenomena in constant discussions, the interpretation of the results obtained tends in some cases to refute or at least to question the real impact of the curvature on the kinetics of the interfaces [ 10 ] or to dispel the 5-parameter description of the reduced mobility [ 76 ] . The observation of interfaces not moving towards their center of curvature ( \left\|\mathbf{v}\right\|/\kappa ratio of opposite sign to the expected one) and/or presenting kinetics very different from other grain boundaries with similar characteristics in the 5-dimensional space indicate that further investigation is required. However, in view of the different proposed equations, the sometimes used recent strong statements that ”GG driving pressure is not correlated to curvature” is inherently ambiguous. Are we speaking of the trace of the curvature tensor (Eq. 2 ) or the curvature tensor itself (Eq. 1 )? In the state of the art, it is ultimately Eq. 2 that is often questioned without taken into account the inclination dependence in experimental or numerical analysis. It is, in fact, both expected and reassuring that Eq. 2 does not hold at the local scale. However, this should not be interpreted as evidence against its validity in simulations involving a large number of interfaces. A major pitfall in recent literature dedicated to GG is probably the growing tendency to consider Eq. 2 as obsolete, not only for describing individual boundaries but even for modeling polycrystalline grain growth in realistic annealing conditions and large systems. This is probably the main symptom emerging from recent literature on the subject, and it likely represents a misconception. Why, in a single-phase system—without second-phase particles and without solute drag at grain boundaries—do the classical models of Burke \& Turnbull [ 12 ] and Hillert [ 76 ] for predicting the mean grain size remain so remarkably accurate in annealing predictions, even though they are fundamentally based on Eq. 2 ? The answer is probably that, at the scale of a few hundred interfaces, Eq. 2 for kinetics-based models, combined with a heterogeneous description of \gamma and an Arrhenius-type temperature dependence for \mu , remains extremely powerful. The same argument could be made for the limiting grain size predicted by the Smith–Zener pinning model [ 73 ] . Finally, it seems important not to dismiss the valuable along with the questionable. Advancing toward Eq. 1 within anisotropic formalisms, not overlooking pressures often considered negligible that may locally become significant during the evolution towards an equilibrium, refining 4D experimental data, and of course improving current models at the GB