AI for science: Progress, challenges, and perspectives

By 2026, artificial intelligence (AI) has completed a qualitative transition from specialized analytical tools to a foundational engine for scientific inquiry, fundamentally altering the paradigms of knowledge generation. In this review, we systematically evaluate the progress of AI for Science (AI4Science), characterized by the shift from text-centric generation to reasoning-oriented intelligence and the deep embedding of foundation models into the entire research lifecycle. We propose a unified, closed-loop research paradigm comprising six core modules: literature research, data analysis, hypothesis formulation, experimental execution, report generation, and autonomous evolution. Furthermore, we synthesize breakthroughs across a broad spectrum of frontier domains—including information science, mathematics, energy, physics, chemistry, materials science, medicine, life sciences, astronomy, geosciences, and social science—highlighting AI’s capacity for overcoming dimensionality constraints and enabling complex inverse design. We then identify critical challenges that persist, such as the scarcity of standardized high-quality data, the interpretability gap, and the physical consistency of generative outputs. Finally, we offer forward-looking perspectives on constructing cross-disciplinary scientific foundation models and fostering a future of human-AI collaborative autonomous research.