Fast prediction of microstructural responses based on realistic material topology is crucial for establishing the process-structure-property chain. This work takes an initial step toward a fully digital framework for developing metallic materials using microscale characteristics. We focus on two objectives: segmenting experimental images to extract microstructural topology and applying a novel operator learning technique to map elastic properties to mechanical deformation and stress states. The operator learning model is trained in a physics-informed, data-free manner which makes the whole process of surrogate modelling very efficient and provide us with more accurate prediction [1]. The segmentation part is performed based on a mask region-based convolutional neural network [2]. We mainly focus on 2D and 3D polycrystalline materials under varying loading scenarios. Compared to solvers such as FEM and FFT, our approach achieves maximum pointwise errors below 5% and average homogenized errors under 1%, while being over 1000 times faster for 3D computations. These results demonstrate the model’s accuracy, computational efficiency, and potential for advanced design applications, with further improvements and extensions discussed.

[1] Rezaei, S., Asl, R.N., Faroughi, S., Asgharzadeh, M., Harandi, A., Koopas, R.N., Laschet, G., Reese, S. and Apel, M., A Finite Operator Learning Technique for Mapping the Elastic Properties of Microstructures to Their Mechanical Deformations. Int J Numer Methods Eng, 2025, 126: e7637. https://doi.org/10.1002/nme.7637.

[2] Viardin, A., Nöth, K., Pickmann, C. et al. Automatic Detection of Dendritic Microstructure Using Computer Vision Deep Learning Models Trained with Phase Field Simulations. Integr Mater Manuf Innov 14, 89–105, 2025, https://doi.org/10.1007/s40192-025-00392-8