Accurate prediction of material properties is crucial for the discovery of novel materials, yet conventional machine learning approaches often struggle with out-of-distribution (OOD) data. This paper introduces Adaptive Manifold Alignment with Spectral Regularization (AMASR), a novel domain adaptation algorithm specifically designed for material property prediction tasks. AMASR integrates manifold alignment techniques with spectral regularization to effectively capture the intrinsic geometric structure of material data while preserving domain-invariant features across source and target domains. Through rigorous mathematical formulation and comprehensive experiments on Matbench datasets, we demonstrate that AMASR significantly outperforms existing domain adaptation methods in OOD scenarios, achieving superior performance in both bandgap prediction and glass classification tasks. Our approach addresses the limitations of current domain adaptation techniques in materials science by explicitly modeling the complex relationships between material compositions and their properties, while maintaining computational efficiency. The proposed algorithm provides a promising direction for enhancing the generalization capability of machine learning models in materials informatics.