Based on a comprehensive analysis of various AI large models and predictive information provided by relevant literature, this paper presents a cautiously optimistic forecast for the development prospects of AI4S over the next five years. After analyzing the automatic generation of top−tier conference papers by The AI Scientist−v2 platform, it is pointed out that there are inherent difficulties in achieving full automation of the research loop, and incomplete verification and drift in research objectives are obstacles that constrain the development of AI4S. From the perspective of computational theory, the root cause of the verification bottleneck in AI4S lies in the semi−decidability of scientific problems, and the boundary of verifiability is the boundary of automation. According to verifiability, scientific and technological problems can be divided into two major categories: R1 problems and R2 problems. The risk of AI4S lies in treating problems that should be governed as R2 problems as R1 problems. This paper sharply points out that AI4S may pose more urgent security risk than AGI, and the risk of AI4S is not "AI becoming smart", but rather "humans losing the right to withdraw ".

AI for Science (AI4S) is reshaping the paradigm of scientific inquiry by integrating scientific priors with data−driven methods. Within this broader shift, AI−driven scientific discovery has become a question of national strategic importance. This paper examines its conceptual boundaries, methodological landscape, and human−AI collaboration mechanisms, and considers its potential implications for the research ecosystem. We first distinguish the two complementary pathways of AI4S—knowledge embedding and knowledge discovery—and the closed loop they form, and argue that the core contributions of knowledge discovery lie in automating inductive reasoning and explicit knowledge formalization. This contribution, however, is concentrated at the level of phenomena and empirical laws; the leap to constructive theories still depends on human mechanistic interpretation and validation. We then propose the Representation–Search–Evaluation (RSE) framework, showing that knowledge−discovery methods differ in the openness of their representation space, with search and evaluation co−evolving around that openness. On this basis, we characterize the core responsibilities that remain with humans in human−AI collaboration: defining epistemic goals, providing semantic constraints, and leading value−based adjudication. Finally, we discuss how knowledge discovery may reshape the research ecosystem. In fields where data density, automation, and validation conditions are relatively mature, the role of scientists may shift from task executors toward a kind of “metacognitive hub” and research organizations may become flatter within groups while growing more differentiated across groups. Such differentiation can be turned into a higher−quality basis for cross−group collaboration through platformized assets, standardized interfaces, and effective validation governance. We emphasize that these ecosystem−level claims are conditional tendencies rather than established facts. The paper’s innovation is to articulate, within the RSE framework, both the methodological trajectory of knowledge discovery and its associated human−AI collaboration mechanism, and on that basis to offer an outlook on the possible changes in research collaboration and organization.

With the rapid development of wide−field surveys, high−cadence monitoring, and multimessenger observations, astronomy is facing a new era of simultaneous growth in data volume, data variety, and observing frequency. At the same time, artificial intelligence is expanding from task−specific models to foundation models and large−language−model−based scientific agents. This article reviews representative advances of these methods in astrophysics over the past five years, with a focus on comparing their functional boundaries in terms of training objectives, data organization, transfer strategies, and scientific roles. The review shows that task−specific models remain the most mature and effective for well−defined problems with stable data distributions; foundation models, pretrained on large−scale heterogeneous data, have demonstrated promising potential for unified representation and cross−task transfer in imaging, spectroscopy, and time−series tasks; scientific agents, however, are still in early exploratory stages, and their roles in literature summarization, tool invocation, and multi−step research workflow orchestration require cautious assessment. The article further examines key challenges, including data quality, generalization, physical interpretability, uncertainty quantification, and reproducibility. It identifies future priorities in building high−quality data infrastructure, establishing cross−survey out−of−distribution validation standards, solidifying software toolchains, and fostering interdisciplinary talent training.

As a forerunner of space applications and the foundation of national space security, space science is entering a new phase characterized by "data explosion, multi−scale coupling, and prominent strategic demands." Traditional research paradigms struggle to address core challenges such as exponential data growth, complex system coupling, autonomous decision−making in deep space exploration, and real−time accurate space weather forecasting. The AI for Science (AI4S) paradigm has emerged as a revolutionary tool for this field, leveraging technologies including deep learning, physics−informed neural networks (PINNs), and causal inference. This paper systematically summarizes the remarkable domestic and international research progress in space science intelligent detection, intelligent recognition, in−depth exploration of mechanisms, and major application practices. It conducts an in−depth analysis of three core challenges: Data infrastructure construction, mechanism−causal modeling, and the implementation of data−intelligent applications. Key solutions such as standardized data governance, physical constraint integration and hybrid modeling, and model light weighting are proposed. Research shows that AI4S has driven a fundamental transformation of space science from "empirical statistics and post−hoc interpretation" to "data−physics collaborative modeling and cognition−driven research." Notable breakthroughs have been achieved globally in on−board intelligent deployment, high−precision planetary landform recognition, and full−chain space weather forecasting. The emergence of several domain−specific large models marks AI4S's entry into a stage of large−scale application in space science. This paper also looks forward to future development trends, laying a solid technological foundation for breakthroughs in space science innovation, support for major space missions, and national space security.

Uncrewed deep-space exploration missions operate under extreme constraints, including vast distances, long communication delays, weak signals, and complex dynamical environments, making conventional ground−control architectures increasingly inadequate for long−duration, complex operations. This review systematically examines advances in intelligent navigation for deep−space exploration. It proposes a perception−decision−execution−auxiliary architecture and synthesizes progress and key bottlenecks in critical technologies, including extreme−environment perception and autonomous cognition, multi−source fusion autonomous navigation, mission planning and decision−making under long−delay conditions, intelligent control in complex dynamical environments, and intelligent navigation auxiliary support technologies for deep space, with particular attention to the integration of artificial intelligence and its future development. Building on this analysis, the review outlines priorities for advancing intelligent navigation in deep space, providing a technical reference for the planning and implementation of China's Phase II planetary exploration program, solar−system boundary exploration, and other future deep−space missions.

In response to the long development cycles and heavy reliance on empirical expertise in materials research, this study elucidates the evolutionary logic underlying the transformation of materials science toward a fourth paradigm of "intelligent autonomous creation" enabled by artificial intelligence (AI). It further analyzes the mechanisms by which cognitive capabilities and decision−making authority are transferred from human researchers to intelligent systems. The paper also outlines the multidimensional pathways through which AI enables materials innovation across three major dimensions: materials design, synthesis planning, and autonomous experimental systems. To address limitations such as data silos, limited physical interpretability, and heavy dependence on upfront computational resources, this work argues that embedding physical mechanisms into statistical models is essential to overcome current capability boundaries. Looking ahead, the deep integration of multimodal foundation models with digital twin technologies is expected to establish an intelligent management framework that covers the entire materials lifecycle. This integration is likely to usher materials research into a self−evolving era characterized by algorithm−driven processes and close human–machine collaboration.

High−throughput computation, high−throughput experimentation, and database development under the Materials Genome initiative have provided the data and platform foundations for artificial intelligence (AI) in materials science. From the interdisciplinary perspective of materials science, mechanics, computational science, automation, and engineering, this article defines materials/mechanics informatics (AI for Materials and Mechanics, AI4M²) and reviews progress in materials design, microstructure characterization, process optimization, mechanics modeling, self−driving experimentation, and scientific agents. The hard cornerstones of AI4M² are AI−enabled experimental robots, high−throughput platforms, intelligent characterization equipment, and self−driving laboratories; its soft cornerstones are AI algorithms, multiscale computation, scientific foundation models, digital twins, and scientific software. Data infrastructure connects the two, while scientific agents link models, tools, experiments, and engineering scenarios. In view of insufficient high−quality data, limited physical interpretability and cross−scale generalization, weak experimental closed loops and indigenous software and equipment, and a shortage of interdisciplinary talent, representative practices in China and abroad are compared. Actionable pathways are proposed, including common data standards and interfaces, shared pilot−scale intelligent laboratories, verifiable indigenous software toolchains, engineering−led demonstration projects, and integrated undergraduate−to−doctoral training.

The rapid development of artificial intelligence is driving a new wave of scientific and technological revolution. In materials science, it is promoting the transformation of data−driven materials R&D systems, represented by Materials Genome Engineering, from an experience−driven model toward an integrated mode combining data−driven approaches with intelligent design. Focusing on the paradigm shift in materials R&D driven by the integration of high−throughput computing and machine learning, this review systematically summarizes the evolution of materials R&D paradigms, the key technological systems of high−throughput computing and machine learning, strategic layouts and progress in platform construction in China and abroad, and analyzes their impacts on the ecosystem of the advanced materials industry, scientific research organization models, and R&D pathways for key materials. High−throughput computing supports materials data generation and large−scale screening, while machine learning facilitates closed−loop iteration in intelligent materials design through rapid pre−screening and knowledge discovery, shifting competition in materials innovation from individual technological capabilities to the integrated capabilities of data resources, computational tools, R&D platforms, and organizational systems. China has initially established a relevant foundation and competitiveness in this field, and future efforts may further focus on improving the autonomy of core software, accumulating high−quality data, refining standard systems, and strengthening the integration of the R&D chain with the industrial chain. The deep integration of high−throughput computing and machine learning is driving the evolution of materials R&D from the fourth paradigm to the fifth paradigm and is continuously reshaping the technological system, industrial ecosystem, and research organization logic of advanced materials innovation.

Drug research and development (R&D) is characterized by long cycles, high costs, and low clinical success rates, and traditional R&D paradigms have struggled to meet the demands of complex diseases and innovative drug development. In recent years, artificial intelligence (AI), leveraging its strengths in multi−source data fusion, complex pattern recognition, and intelligent decision−making, has been driving a paradigm shift in drug R&D from experience−driven to data−driven and knowledge−driven approaches. This article systematically reviews the latest advances in AI−empowered drug R&D worldwide, focusing on key applications across the full drug R&D pipeline and technological development trends, and synthesizes representative studies to summarize AI's vital roles in improving R&D efficiency, optimizing decision−making, and fostering paradigm changes. The review shows that AI has become an essential bridge linking biomedical data, computational models, and experimental validation. It has not only significantly enhanced the efficiency of early−stage target discovery, molecular design, and drug evaluation, but has also spurred the development of novel R&D modalities such as automated experimentation, self−driving laboratories, and virtual cells, thereby providing new technical pathways toward an intelligent drug R&D system. However, the broad application of AI in drug R&D remains constrained by insufficient high−quality data, limited model interpretability and generalizability, difficulties in multimodal knowledge fusion, and the lack of robust experimental validation and standardized evaluation frameworks; its clinical translational value thus requires further verification. Looking ahead, efforts should be directed toward strengthening high−quality data resources and sharing systems, developing interpretable AI models that integrate biological mechanisms and physical laws, promoting deep integration of generative AI, multi−agent systems, and self−driving laboratories, refining model evaluation standards and regulatory frameworks, and fostering collaborative innovation between AI and experimental sciences, so as to accelerate the evolution of AI−empowered drug R&D toward intelligent, automated, and closed−loop paradigms.

Bio−inspired navigation is an important embodiment of bio−inspired intelligence that empowers autonomous perception and efficient navigation in unmanned systems. By mimicking the processes of biological systems in perceiving, acquiring, integrating, and utilizing spatial, environmental, and target−related information, bio−inspired navigation provides new theoretical insights and technical approaches for overcoming the bottlenecks of autonomous navigation in interference and adversarial environments. It has become a frontier research direction involving the interdisciplinary integration of information science, bionics, and instrumentation engineering. Starting from the concepts and functions of bio−inspired intelligent systems, this paper elaborates on the research connotation and architectural framework of bio−inspired navigation. The research progress of representative bio−inspired navigation technologies, including inertial navigation, visual navigation, polarization navigation, and geomagnetic navigation, is systematically reviewed. The bio−inspired mechanisms, information perception mechanisms, and technological implementation forms of different navigation approaches are analyzed in detail. Furthermore, the future development trends of bio−inspired navigation are discussed from three perspectives: neural−enabled intelligent motion information estimation (method), intelligent navigation hardware based on bio−like organ collaborative perception (system), and bio−inspired intelligent navigation with adaptive capabilities against uncertain interference (behavior). This review provides references for the development of next−generation autonomous navigation technologies.

In recent years, the rapid development of next−generation information technologies has propelled artificial intelligence (AI) from a versatile technical tool to a strategic cornerstone driving the emergence of new productive forces and reshaping national technological competitiveness. AI's role is undergoing a profound shift—from an "enabler" to a "definer"—fundamentally transforming research paradigms and trajectories in interdisciplinary fields. This paper systematically traces the five evolutionary stages of scientific research paradigms, focusing on the fifth: AI for Science (AI4S). Centered on AI−driven innovation, AI4S integrates advanced technologies (e.g., convolutional neural networks, graph neural networks, diffusion models, large models) to systematically reconstruct scientific logics and technological pathways in frontier disciplines, including biomedicine, humanoid robotics, and quantum science. By doing so, it realizes intelligent transformation across the entire research lifecycle. The study underscores that authentic AI4S embodies a synergistic fusion of human creativity and scientific exploration. By automating repetitive tasks (e.g., low−level reasoning, experimental design), AI empowers researchers to concentrate on high−level problem formulation and strategic scientific planning, accelerating discoveries and breakthroughs. Human ingenuity remains the bedrock of science; free exploration and original theorization are indispensable. Through human−AI collaboration, AI4S propels scientific research toward higher−order intelligence and interdisciplinary integration, defining the future of technological advancement.

To address the growing demand for thermal dissipation substrates in high−power electronic devices, this study focuses on overcoming the mechanical limitations of aluminum nitride (AlN) ceramics and the reduction in thermal conductivity associated with conventional toughening methods. A phase−compatible toughening strategy using one−dimensional AlN whiskers is proposed. AlN whiskers were synthesized via direct nitridation and incorporated into AlN ceramics through dry pressing and atmospheric pressure sintering at elevated temperature. The results indicate that the AlN whiskers form strong interfacial bonds with the matrix, effectively reducing thermal boundary resistance. Its high aspect ratio creates continuous, highly efficient heat conduction pathways, which collectively enhance the thermal and mechanical properties of the AlN ceramic. The optimized composites achieved a thermal conductivity of 189.49 W·m⁻¹·K⁻¹, a flexural strength of 255.13 MPa, and a fracture toughness of 3.25 MPa·m^1/2, corresponding to enhancements of 7.2%, 150.8%, and 24.5%, respectively, compared to unreinforced AlN. Microstructural analysis combined with finite element simulations confirms that the AlN whiskers effectively redistribute mechanical stress and suppress crack propagation. This work provides a viable approach for developing high−performance ceramic substrates with integrated thermal management and structural reliability for advanced electronic packaging applications.

As the demand for nuclear medicine diagnosis and treatment increases due to the population aging, a stable and reliable supply of medical isotopes, which serve as the material foundation for nuclear medicine, is a crucial prerequisite for promoting the development of nuclear medicine and safeguarding human health. This paper summarizes the current states of reactor−produced medical isotope production and supply, noting that China remains highly dependent on imports for the medical isotopes commonly used in clinical diagnostics and therapy. To address this, China has launched new projects to construct medical isotope production reactors. It is projected that China’s supply of key medical isotopes will continue to rely primarily on imports until 2027; after 2027, the self−sufficiency capacity for certain isotopes will significantly improve, and in the long term, it is expected to achieve complete self−sufficiency and even partial exports. However, the realization of the expected production capacity following the commissioning of these facilities still face multiple risks and challenges, such as the immaturity of production reactor operation management and supporting infrastructure, unclear pricing mechanisms for reactor−produced medical isotopes, and the incomplete spatial layout and commercial structure of research reactors and processing plants, which may lead to supply fluctuations or even shortages. It is proposed to strengthen the top−level design of medical isotope supply security, coordinate the planning, layout, and construction progress of production facilities; improve the operational management and supporting infrastructure of medical isotope production reactors, establish a pricing model and commercial structure for medical isotopes that aligns with national conditions, to promote the sustainable development of the supply chain.

Federated learning, as a distributed machine learning paradigm, enables collaborative multi-party learning while preserving data privacy, and has been widely deployed in various critical domains. However, its distributed nature also renders it highly vulnerable to backdoor attacks, posing serious security threats. This paper first provides a detailed analysis of the methodological taxonomies and evolutionary directions of backdoor attacks and defenses in the field of federated learning, and summarizes the latest research advances in related areas. Backdoor attack research is shifting toward realistic strategies that simultaneously pursue stealth, persistence, and low data dependency, giving rise to a variety of novel attack vectors. Meanwhile, backdoor defenses are increasingly pursuing greater universality against diverse attack types, as well as a balance between computational overhead and defensive strength within the federated learning environment. Finally, this paper reveals that backdoor attacks and defenses in federated learning are facing increasingly complex challenges, and calls for a proactive shift toward more systematic and integrated designs. On the attack side, research should pivot toward multi-objective balanced attack strategies; on the defense side, emphasis should be placed on intelligent and adaptive mechanisms, while integrating technologies such as privacy-preserving computing to strengthen security guarantees, thereby facilitating its deployment in real-world federated learning environments.

Artificial intelligence has become a strategic commanding height in global technological competition, with China and the United States exhibiting significant differences in development pathways, institutional design, and technological innovation. This paper constructs a three−dimensional analytical framework of "policy, technology, and scientific research" to systematically compare the similarities and differences in AI development between China and the United States, and for the first time incorporates the token economy as a comparative dimension, revealing new forms of AI commercialization. The study draws on the latest policy texts, including China's Opinions on Deepening the Implementation of the "AI+" Initiative, the 15th Five−Year Plan, and the U.S. Winning the Race: America's Action Plan for Artificial Intelligence, combined with frontier model technological breakthroughs since 2025 and data from Stanford's 2026 AI Index Report. The findings indicate that China has formed a development model dominated by industrial implementation and scenario−driven feedback, with policies characterized by vertical, phased planning; whereas the United States has formed a development model led by breakthroughs in artificial general intelligence, with policies characterized by itemized, horizontal execution. In terms of scientific research output, the two countries exhibit a pattern where "China leads in quantity, while the United States leads in quality," while the gap in frontier model performance has significantly narrowed. On this basis, the paper proposes three systematic recommendations for improving China's AI governance system, namely, constructing executable policy mechanisms, establishing a multi−dimensional quantitative evaluation system, and improving mechanisms for incorporating technical experts into decision−making. It also advocates maintaining the strength of an "application−oriented" approach while strengthening forward−looking deployments in artificial general intelligence (AGI) and embodied intelligence, so as to form a dual−wheel driven development paradigm characterized by "application advantages+foundational breakthroughs".

The year 2026 marks the centenary of the birth of Mr. Yu Min, a distinguished scientist who made outstanding contributions to China's "two bombs and one satellite" program and a recipient of the Medal of the Republic. As one of the leading figures in China's nuclear weapons theoretical research and advanced defense technology development, and as a key contributor to the breakthrough in hydrogen bomb principles, Yu Min, together with Peng Huanwu, Zhu Guangya, Deng Jiaxian, and other scientists, pioneered an independent path for nuclear weapons development under extremely difficult conditions. Drawing on firsthand accounts from those who worked with him, and on the basis of a systematic review of Yu Min's scientific career, this paper further distills and elucidates the generative logic, specific manifestations, and practical value of the six dimensions of his scientific spirit—patriotism, innovation, pragmatism, dedication, collaboration, and education. Yu Min's scientific spirit embodies both philosophical depth and personal inspiration. Against the strategic backdrop of the Fourth Plenary Session of the 20th Central Committee of the Communist Party of China, which set the goal of "achieving high−level scientific and technological self−reliance and building a world leader in science and technology," studying and promoting this spirit holds urgent and profound significance for our times. It inspires scientific and technological workers to keep the country's overall interests at heart, boldly venture into uncharted innovation frontiers, uphold truth−seeking and pragmatism, devote themselves selflessly, enhance collaborative efficiency, and cultivate a contingent of strategic scientists.