Abstract: With the rise of artificial intelligence for science (AI4S), attributed graph optimization has gradually become a pivotal link connecting graph machine learning with emerging fields such as biomedicine and new materials, demonstrating broad application potential. Traditional optimization methods lack domain knowledge integration, and suffer from inefficiency in actual scenarios, low efficiency of black-box evaluation, and insufficient controllability of the optimization process. In this article, we systematically review the cutting-edge technologies in attributed graph optimization for AI4S. First, this article deeply analyzes the modeling and representation methods of attributed graphs, exploring how to improve model adaptability to specific scientific tasks through accurate graph representation mechanisms. Furthermore, we parse the basic principles of black-box optimization and deep surrogate models, discussing how to achieve efficient approximation of the black-box evaluation process to enhance the overall efficiency and accuracy of the model. Finally, we focus on the mechanisms of large language models (LLMs) in domain knowledge injection and decision support, aiming to enhance the interpretability and controllability of the optimization process. Research on attributed graph optimization for AI4S not only helps promote the deep interdisciplinary integration of computer science and other disciplines but also accelerates the practical application of graph machine learning technologies in solving real-world scientific problems across multiple fields, thereby creating greater economic and social benefits.