In this paper, we propose a deep-learning algorithm, Data-driven simulation assisted Physics-learned Artificial Intelligence (DPAI), to simulate heat diffusion in
large-grain polycrystalline material. The DPAI model utilizes an encoder-decoder
architecture with convolutional long short-term memory (ConvLSTM), which captures
the spatio-temporal representation from input sequences. The DPAI model learns the
physics of heat diffusion in the material from training datasets. This model is trained
with Finite Element (FE) simulation datasets consisting of a varying number of grains
in a microstructure of polycrystalline material with a single-point heat source at the
center. The arbitrary plane of the 3D microstructure of these materials is represented
using 2D Voronoi tessellations. These configurations are used to model the geometry
of the 2D Computer-Aided Design (CAD) model. Each cell of the Voronoi tessellation
represents one grain of the microstructure. This CAD model is used as the input to
the FE solver for solving heat diffusion equations. To model the near-realistic material
anisotropy and accurately measure temperature differences at cell boundaries, a smaller
mesh size is required in FE modeling, which takes considerable solver time. Therefore,
the proposed model significantly reduces the computational time while maintaining
reasonable accuracy compared to conventional FE simulation. The effectiveness of
the trained DPAI model is examined by modeling larger domain problems involving a
greater number of grains and varying material properties.