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

TL;DR

This paper introduces an advanced deep learning model that predicts microstructure evolution during complex thermal treatments by incorporating thermal profile modulation, significantly improving accuracy while maintaining computational efficiency.

Contribution

The study extends previous ConvLSTM-based grain growth models by integrating FiLM for thermal conditioning, enabling accurate predictions under variable thermal profiles.

Findings

Achieved SSIM of up to 0.93 in test scenarios.

Maintained inference time of seconds per prediction.

Reduced grain size prediction error below 3.2%.

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 kelvin per second to 10 kelvin per second. 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 and mean grain size error below 3.2%. Despite the architectural extensions, inference time remains on the order of seconds per prediction sequence, preserving the computational advantage over PDE-based simulations.