In the traditional micromechanics context, the size and volume fraction of inclusions in heterogeneous composites are considered the primary microstructural parameters for controlling macroscopic properties, rather than other factors. Challenges persist in accurately predicting effective properties due to a limited understanding of extensive multi-scale microstructural data using micromechanics-based methods such as homogenization and finite element analysis. In this research, both supervised and unsupervised learning approaches are employed to explore the limits of artificial intelligence comprehension. In the case of supervised learning, analysis of scanning electron microscope (SEM) images of nickel-based superalloys from a high-throughput experiment involves defining 23 microstructural descriptors. Subsequently, 10 descriptors are selected to reduce the computational cost of the deep neural network (DNN) with the support of the shallow neural network (SNN). Additionally, in order to enhance DNN accuracy, new training sets are proposed by incorporating these 10 descriptors along with two additional ones: area distribution and one heat treatment parameter - cooling rate. In conclusion, it is demonstrated that the supervised learning approach surpasses the predictive capabilities of existing physics-based constitutive models. In the case of unsupervised learning, the prediction of the effective thermal conductivity of thermal insulation composite materials is achieved through the utilization of a convolutional neural network (CNN). The CNN model is trained and validated using microstructural images generated from numerical approaches as input data and the conductivities predicted by the Lattice Boltzmann Method (LBM) as output data. Subsequently, the CNN predictions are compared to experimental results as well as those obtained from analytical and computational micromechanics methods. It is worth noting that the proposed CNN model demonstrates accurate prediction of the thermal conductivity for materials featuring novel microstructures that were not part of the training set. In summary, harnessing artificial intelligence to capture the scattering characteristics of heterogeneous materials enables both DNN and CNN models to achieve more efficient predictions compared to traditional methods. This highlights the potential of machine learning in advancing materials science and expediting the development of materials with desired properties.