Advances in AI, especially deep learning, have enabled new approaches to the classic ill-posed inverse problem of heterogeneous material parameter identification, opening new pathways for its solution. To address these shortcomings of conventional inverse identification methods—namely, poor noise resistance and low computational efficiency in heterogeneous material inversion—the application of artificial intelligence is explored by investigating both supervised and unsupervised learning paradigms. It is revealed that deep learning methods can effectively capture the complex correspondence between material parameters and mechanical responses by leveraging the nonlinear mapping capability in feature space. By using finite element simulation to generate training data and applying a Conditional Generative Adversarial Nets (CGAN)for reconstruction, the data-driven supervised learning approach yields better reconstruction of heterogeneous material distributions than traditional optimization algorithms. However, acquiring authentic and effective training datasets is often highly challenging. Therefore, we propose an unsupervised learning method based on the Deep Image Prior (DIP), which leverages the implicit regularization property of neural networks to accomplish the reconstruction of material distributions in the target domain without requiring any training data. Even without pretraining the network model and without utilizing any training data, the method yields favorable reconstruction outcomes for heterogeneous materials. In summary, the inverse identification of heterogeneous materials is implemented using both supervised and unsupervised learning approaches. Artificial intelligence technology exhibits significant potential in enhancing both the speed and accuracy of reconstructing material distributions within target domains.