Focus on the convergence of artificial intelligence (AI) innovation in scientific research.

 

AI for Science (AI4S) represents the convergence of artificial intelligence (AI) innovation in scientific research and AI-driven scientific discovery, demonstrating their deep integration1, and the establishment of a transformative research paradigm.


Traditional research paradigms can be categorized as empirical induction (experimental science), theoretical modeling (theoretical science), computational simulation (computational science), and data-intensive science2. The experimental scientific paradigm generates empirical laws from observations of natural phenomena and reproducible experiments, but does not provide the theoretical foundations that would explain these laws at a fundamental level.

The theoretical paradigm also begins with observations of natural phenomena and reproducible experiments. From these it identifies fundamental scientific problems, and formulates formal hypotheses, and ultimately develops theories through systematic logical reasoning and mathematical analysis. However, verifying these theories within complex systems remains a significant challenge.

Computational science employs numerical methods to simulate complex systems based on scientific models. However, it must simplify these models and requires high-precision computation, inherently limiting fidelity and efficiency.

With technological advances and the exponential growth of data, a new research paradigm of data-intensive science has emerged, using data mining techniques to automatically identify statistical patterns from large-scale datasets, reducing reliance on priori scientific hypotheses. However, it faces limitations in establishing causal relationships, processing noisy or incomplete data, and discovering principles in complex systems.

Modern research confronts complexity challenges, in which interconnected natural, technological, and human systems exhibit multi-scale dynamics across time and space1. Traditional research methods struggle to address these complex challenges effectively, demanding new methods. The need to establish causality has driven the development of innovative inference methodologies capable of handling modern data challenges.

To address the scarcity of high-quality scientific data and other problems, generative AI technologies such as diffusion models and large language models (LLMs) have been developed. For overcoming limitations in complex system modeling, knowledge-guided deep learning approaches that embed prior knowledge into deep neural networks have been established, significantly enhancing generalization and improving interpretability, such as physics-informed neural networks3.

AI innovation is reshaping traditional research processes and accelerating discovery. AI integrates data-driven modeling with prior knowledge, which is called model-driven, automating hypothesis generation and validation, enabling autonomous and intelligent experimentation, and promoting cross-disciplinary collaboration. Traditional scientific discovery centres on experimental observations and theoretical modeling, formulates scientific hypotheses and induces general principles, such as physical laws. In contrast, AI employs a model-driven approach to automatically discover hidden patterns from large-scale data, circumventing the need for hypotheses.

Traditional scientific discovery involves generating and validating candidate hypotheses from a large solution space, often characterized by low efficiency and challenges in identifying high-quality solutions4. AI harnesses its powerful data processing and analytical capabilities to navigate solution spaces more efficiently, enabling the generation of high-quality candidate hypotheses. For instance, machine learning can assist mathematicians in uncovering new conjectures and theorems5.

Scientific research depends on the experimental validation of theories. Traditional approaches to experimental design and optimization often rely on manual expertise and iterative trial-and-error processes, which are expensive and inefficient. This is particularly evident in fields such as materials synthesis and fusion experiments.

The integration of AI and robotics can facilitate automated experimental design and execution, leveraging real-time data to refine parameters and optimize both experimental workflows and candidates. AI excels at integrating data and knowledge across fields, breaking down academic barriers and enabling deep interdisciplinary integration to tackle fundamental challenges. This cross-disciplinary collaboration has not only pushed the boundaries of research, but given rise to emerging disciplines, such as computational biology, quantum machine learning, and digital humanities.

Looking to the future, key challenges in AI4S include improving cross-scale modeling, enhancing AI generalization in data-scarce fields, and pushing the boundaries of AI-assisted hypothesis generation. Future breakthroughs may come from interdisciplinary knowledge graphs, reinforcement learning-driven closed-loop systems, and interactive AI interfaces that refine scientific theories.

The rapid advancement of AI4S signifies a profound transformation: AI is no longer just a scientific tool but a meta-technology that redefines the very paradigm of discovery, unlocking new frontiers in human scientific exploration.

To read the full AI for Science 2025 report.