Predicting Grain Growth Evolution Under Complex Thermal Profiles with Deep Learning through Thermal Descriptor Modulation

Pungponhavoan Tep, Marc Bernacki · 2026

Read Predicting Grain Growth Evolution Under Complex Thermal Profiles with Deep Learning through Thermal Descriptor Modulation by Pungponhavoan Tep, Marc Bernacki (2026) 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-2603-24090.txt.

Introduction

Tep and Bernacki extend a deep learning model that predicts grain growth under changing heat profiles. Their earlier ConvLSTM model ran only at constant temperature or a single heating rate. Real heat treatment varies over time, and grain boundary motion depends on that history. The new model adds feature-wise linear modulation for thermal conditioning. It trains on profiles with rates from 0.01 to 10 kelvin per second. Across three test cases, image similarity reaches 0.93 and mean grain size error stays below 3.2 percent.

Catalog record

CREATORPungponhavoan Tep, Marc Bernacki
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE30126
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.24090

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 Evolution Under Complex Thermal Profiles with Deep Learning through Thermal Descriptor Modulation Pungponhavoan Tep, Marc Bernacki Abstract Predicting microstructure evolution during thermomechanical treatment is essential for determining the final mechanical properties of a material, yet conventional simulations based on Partial Differential Equations (PDEs) remain computationally expensive. Our prior Deep Learning (DL) framework using Convolutional Long Short-Term Memory (ConvLSTM) has proven effective in accelerating grain growth prediction, though its applicability was limited to constant-temperature or single-rate thermal profiles. As the model was trained exclusively under constant thermal conditions, it cannot account for the thermal history dependence of grain boundary kinetics, fundamentally limiting its applicability to the time-varying thermal profiles characteristic of industrial heat treatment processes. This study extends the previous framework by incorporating Feature-wise Linear Modulation (FiLM) for thermal conditioning to predict grain growth under complex, time-varying thermal profiles. The model was trained on a large dataset of grain growth evolution under thermal profiles with heating and cooling rates ranging from 0.01 K s − 1 0.01\text{\,}\mathrm{K}\text{\,}{\mathrm{s}}^{-1} to 10 K s − 1 10\text{\,}\mathrm{K}\text{\,}{\mathrm{s}}^{-1} . The results demonstrate that the proposed thermal conditioning mechanism enables the model to capture the influence of variable thermal profiles on grain boundary migration kinetics. Across the three test scenarios of increasing complexity, the model achieved a Structural Similarity Index Measure (SSIM) of up to 0.93 0.93 and mean grain size error ( R ¯ \overline{R} error) below 3.2 % 3.2\text{\,}\mathrm{\%} . Despite the architectural extensions, inference time remains on the order of seconds per prediction sequence, preserving the computational advantage over PDE-based simulations. 1 Introduction The evolution of microstructure during thermal processing determines the final mechanical properties of polycrystalline materials. Among relevant mechanisms, grain growth is driven by the reduction of grain boundary energy and governs the microstructural outcome of heat treatment processes [ 20 , 8 ] . Conventional Partial Differential Equation (PDE)-based simulations, including front-tracking [ 5 , 7 , 6 , 16 , 17 ] , phase-field [ 24 , 15 , 11 , 4 , 21 ] , and level-set methods [ 2 , 13 ] , provide accurate descriptions of microstructure evolution but incur substantial computational cost due to their sequential timestep calculations. Our previous study [ 22 ] addressed this limitation through a Deep Learning (DL) framework based on an encoder-decoder architecture coupled with Convolutional Long Short-Term Memory (ConvLSTM), achieving up to 90\times speedup over our in-house front-tracking simulation ToRealMotion (TRM) [ 5 , 7 ] while maintaining high-fidelity predictions, with a mean grain size error ( \overline{R} error) of 0.07\text{\,}\mathrm{\%} and a Structural Similarity Index Measure (SSIM) of 86.71\text{\,}\mathrm{\%} under isothermal grain growth conditions. However, that framework was trained and evaluated exclusively under constant thermal conditions. In practice, industrial heat treatment involves complex sequences of heating, holding, and cooling stages with varying thermal rates. Since grain boundary migration kinetics depend on temperature through Arrhenius-type mobility [ 3 ] , an identical microstructure can evolve along different trajectories depending on the imposed thermal profile, which the previous model could not capture without explicit thermal descriptors. Accordingly, this study extends the previous DL framework to non-isothermal grain growth prediction through Feature-wise Linear Modulation (FiLM) [ 19 ] thermal conditioning. By providing instantaneous temperature ( T ) and thermal rate ( \mathrm{d}T/\mathrm{d}t ) as explicit conditioning inputs alongside microstructure images, the model adapts its predictions to varying thermal conditions while preserving the computational efficiency of the original architecture. Performance was evaluated across three test scenarios of increasing complexity, from a simple heating-holding-cooling cycle to a complex multi-cycle thermal profile excluded from training, to assess both prediction accuracy and generalization capability. 2 Methodology 2.1 Model Architecture The framework proposed in the previous study, as illustrated in Figure 1 (a), consists of a Convolutional Autoencoder coupled with ConvLSTM, which compress microstructure images into a latent space, model spatiotemporal grain growth patterns, and reconstruct full-resolution predictions autoregressively. In that framework, only sequences of microstructure images were accepted as input, with boundary migration kinetics implicitly learned from constant-rate training data. (a) (b) Figure 1: Neural network architectures for grain growth prediction: (a) encoder-decoder + ConvLSTM framework from the previous study; (b) extended framework with thermal conditioning through FiLM, where a thermal conditioning module generates scale ( \gamma ) and shift ( \beta ) parameters to modulate the latent features based on instantaneous T and \mathrm{d}T/\mathrm{d}t . To extend prediction capability to complex thermal profiles, the architecture is augmented with a thermal conditioning mechanism through FiLM [ 19 ] , as illustrated in Figure 1 (b). At each timestep, T and \mathrm{d}T/\mathrm{d}t are processed through a Multi-Layer Perceptron (MLP) to generate scale ( \gamma ) and shift ( \beta ) parameters that modulate the latent features according to: \mathbf{z}^{\prime}=\gamma\odot\mathbf{z}+\beta, (1) where \mathbf{z} and \mathbf{z}^{\prime} are the original and modulated latent representations, respectively. This element-wise affine modulation allows the model to adapt grain boundary migration kinetics dynamically in response to instantaneous thermal conditions throughout both input encoding and autoregressive prediction, without modifying the core encoder-decoder architecture. 2.2 Dataset Generation The identical 18 initial microstructure states used in the previous study were retained, spanning a range of mean grain sizes, domain sizes, and topological characteristics representative of polycrystalline microstructures, as listed in Table 1 . For each initial state, grain growth evolutions were simulated using the in-house TRM [ 5 , 7 ] under different thermal profiles of heating, isothermal holding, and cooling phases at thermal rates between 0.01\text{\,}\mathrm{K}\text{\,}{\mathrm{s}}^{-1} and 10\text{\,}\mathrm{K}\text{\,}{\mathrm{s}}^{-1} , as shown in Figure 2 and listed in Table 2 . The grain boundary mobility, \mu , is defined through an Arrhenius law: \mu=\mu_{0}\exp\left(-\frac{Q}{RT}\right), (2) with \mu_{0} a constant pre-exponential term, Q the activation energy, R the gaz constant and T the absolute temperature. The grain boundary energy \gamma is assumed constant. The physical values used for \mu_{0},\ Q,\ \text{and}\ \gamma are also summarized in Table 2 and are representative of the 304L stainless steel [ 14 ] . Each sequence spanned 59\text{\,}\mathrm{min} with microstructure states captured at 1\text{\,}\mathrm{min} intervals alongside the corresponding T and \mathrm{d}T/\mathrm{d}t values computed as backward finite differences. After applying the same data augmentation techniques as the previous study, the final dataset contained 6727 evolution sequences, partitioned into training, validation, and test sets at 80\text{\,}\mathrm{\%} , 10\text{\,}\mathrm{\%} , and 10\text{\,}\mathrm{\%} , respectively, using a stratified split by domain size. This stratification ensured that each domain size was proportionally represented across all three subsets, ensuring unbiased learning and evaluation across different microstructural scales. Parameter Value(s) Description a ( \mathrm{mm} ) 2 , 3 , 4 , 5 Side length of the square domain. Distribution \mathcal{N}(\overline{R},\,\sigma^{2}) The type of distribution used for the ECR distribution (Laguerre-Voronoï tessellation [ 1 , 9 ] ). \overline{R} ($\text{\,}\mathrm{\SIUnitSymbolMicro m}$) 20 The mean value of the ECR distribution. \sigma ($\text{\,}\mathrm{\SIUnitSymbolMicro m}$) 2 , 4 , 8 , 16 , 32 The standard deviation of the ECR distribution. Table 1: Parameters for generating initial microstructure states using LavoGen. Equivalent Circle Radius (ECR) = the radius of the circle having the same area as the considered grain (in \text{\,}\mathrm{mm} or \text{\,}\mathrm{\SIUnitSymbolMicro m} ). Parameter Value(s) Description T ($\text{\,}\mathrm{K}$) 293.151325.15 The temperature range. \mathrm{d}T/\mathrm{d}t heating ($\text{\,}\mathrm{K}\text{\,}{\mathrm{s}}^{-1}$) 0.11 The heating rate range. |\mathrm{d}T/\mathrm{d}t| cooling ($\text{\,}\mathrm{K}\text{\,}{\mathrm{s}}^{-1}$) 0.0110 The cooling rate

Read the complete text