When Models Become a New Engine for Scientific Research: AI for Science Ushering in a New Cycle of Scientific Discovery

AlphaFold2 achieved high‑accuracy protein‑structure prediction at the CASP14 competition in 2020 and was formally published in Nature in 2021. This breakthrough brought AI‑for‑Science to the attention of the global research community. Today, a wide array of AI‑powered research products keep iterating, with models deeply embedded in research workflows such as candidate‑molecule screening, simulation, and data interpretation. Yet a critical question remains: are models merely efficiency‑boosting tools, or are they emerging as entirely new engines driving scientific discovery? To grasp the true nature of this new scientific cycle, we need to take a long‑term historical perspective and situate it within the evolution of successive scientific‑research paradigms.

1. The Long Trajectory of Paradigm Shifts: From Hypothesis‑Driven to Data‑Driven Scientific Traditions

A scientific paradigm denotes the dominant set of working logics for posing, investigating, and validating scientific questions in a given era. The birth of a new paradigm rarely overturns existing wholesale systems; Instead, it expands human capacity to understand the objective world. Iterations across academic research platforms essentially represent adaptations to such paradigm shifts.

1.1 Operating Logic and Inherent Limitations of the Hypothesis‑Driven Paradigm

For a long time, the hypothesis‑driven research paradigm has constituted the mainstream working model for most basic research. Building on established theories and observational phenomena, researchers formulate scientific hypotheses, then design and conduct experiments or observations to verify or falsify those hypotheses, completing one iteration of scientific understanding.

This approach relies heavily on researchers’ theoretical grounding and scientific intuition. For instance, many early gene‑function studies in molecular biology began with functional conjectures put forward by researchers, followed by controlled experiments for validation.

Nevertheless, the hypothesis‑driven paradigm has inherent bottlenecks. Human‑formulated hypotheses are bounded by individual knowledge and imagination. When confronting high‑dimensional, multi‑variable complex systems, humans cannot exhaust all possible combinatorial conjectures, meaning promising research paths may be inadvertently overlooked.

1.2 Rise and Practical Shortcomings of the Data‑Driven (Fourth) Paradigm

In 2007, computer scientist Jim Gray presented the concept of the Fourth Paradigm in a keynote speech: after experiment, theory, and simulation, scientific research had entered an era of data‑intensive science, where patterns could be mined directly from large‑scale datasets without full reliance on pre‑established hypotheses. The speech was later compiled into the 2009 book The Fourth Paradigm: Data‑Intensive Scientific Discovery.

Climate modelling, multi‑omics big data, and astronomical sky surveys are typical real‑world applications of the Fourth Paradigm. In astronomy, for example, sky‑survey telescopes continuously generate terabyte‑scale observational datasets. Rather than setting predefined observation targets, researchers leverage data‑mining algorithms to identify previously unknown celestial objects from massive astronomical imagery.

Still, the data‑driven paradigm carries practical drawbacks. It hinges on high‑quality, sufficiently large annotated datasets. Where high‑quality domain‑specific data is scarce or suffers from sampling bias, patterns extracted from data can prove misleading. Furthermore, mere statistical correlations derived from data cannot automatically yield complete causal explanations; Correlation does not equate to scientific mechanisms.

Each successive research paradigm addresses weaknesses of its predecessor, yet none eliminates the need for human reasoning and interpretation in scientific work.

By understanding the respective strengths and limitations of hypothesis‑driven and data‑driven paradigms, we can objectively assess the innovations brought by the model‑driven approach embodied in AI‑for‑Science.

2. Rise of Model‑Driven Research: Origins and Paradigm Debates Surrounding AI‑for‑Science

The 2020s have witnessed the confluence of expanding computing power, growing repositories of high‑quality research datasets, and rapid maturation of deep‑learning model engineering. Combined, these factors have moved AI‑for‑Science from sporadic explorations toward large‑scale implementation. Many research tools integrate large‑model capabilities into day‑to‑day research workflows. UniResearch, for example, develops solutions for fundamental research tasks including literature processing and information organization.

2.1 Landmark Milestones for the Model‑Driven Paradigm

The success of AlphaFold2 popularized the notion of “model‑driven research”. Even so, the academic community has yet to reach consensus on its status as a research paradigm.

2.2 Two Core Scholarly Perspectives

• Perspective 1: Model‑Driven Research Represents the Fifth Scientific Paradigm Some scholars argue that whereas the Fourth Paradigm focuses on extracting statistical patterns from existing data, the model‑driven paradigm learns inherent scientific laws to proactively generate candidate entities and run virtual simulations. It produces previously unobserved candidate solutions for subsequent experimental validation. Acting as a generator for scientific hypotheses and virtual experiments, the model enables a complete new research workflow and therefore qualifies as an independent paradigm.
• Perspective 2: Model‑Driven Research Is a Technical Enhancement of the Fourth Paradigm Other researchers counter that model‑driven approaches constitute only a technical enhancement rather than an entirely new paradigm. They note that model training still draws heavily on large‑scale research datasets and remains grounded in data‑intensive science. All outputs generated by models ultimately require real‑world experimental validation. Thus, model‑driven work stays within the broader framework of data‑driven research, merely raising the upper bound for data processing and generation.

2.3 Essence of the Debate: Divergent Criteria for Defining Paradigms

This academic dispute yields an important insight: paradigm classification largely depends on evaluation criteria. If judged by tooling and algorithm forms, model‑driven research demonstrates distinctive features. If evaluated against the underlying research workflow — data input → analysis → experimental validation — it fits within the Fourth‑Paradigm framework.

The model‑driven paradigm did not emerge through sudden revolution; it evolved upon foundations laid by earlier generations of scientific paradigms.

Academic discussions over its paradigmatic status continue. Meanwhile, China, the United States, and Europe have rolled out top‑level strategies to facilitate AI‑for‑Science deployment. The following section compares major global economies and their distinct development pathways shaped by different resource endowments.

3. Global Landscape: A Comparison of Top‑Level Strategies and Priorities for AI‑for‑Science across China, the US and Europe

Faced with the new research cycle ushered in by model‑driven science, major global economies have released dedicated strategic documents for AI‑for‑Science. Due to differences in industrial foundations, research systems, and resource endowments, their development priorities diverge noticeably.

3.1 China: Balancing Foundation Models, Infrastructure, and Industrial Applications

In August 2025, the State Council issued Opinions on Further Implementing the “AI +” Initiative, which ranks “AI + Science and Technology” as its primary priority action. The document explicitly calls for AI‑driven shifts in research paradigms and accelerated “from‑zero‑to‑one” scientific discoveries. China is simultaneously building research datasets and computing infrastructure, promoting industry‑university‑research collaborative initiatives, and pursuing applied research on key scientific problems in materials, biomedicine, energy, and other strategically important industries.

Against the backdrop of digital transformation in research, digital laboratories and collaborative research platforms form critical components of China’s research‑system development. Domestic research tools including UniResearch are adapting their products to support these priorities.

3.2 The United States: Dual‑Driven by Industry and Public Research Institutions

The US follows a dual‑track development model led by private‑sector enterprises alongside public research organizations. Leading technology companies undertake much of the work on general‑purpose scientific foundation models, while national research agencies and universities focus on defining scientific questions, conducting experimental validation, and pursuing fundamental‑science exploration. A 2025 report by the US National Academies of Sciences on foundation models for energy research identifies gaps in computing resources, high‑quality annotated datasets, and interdisciplinary talent. It also flags reproducibility risks stemming from proprietary closed‑source commercial models in research.

3.3 European Union: Built on Open Science, Prioritizing Public Research Infrastructure

On October 8, 2025, the European Commission published the European Strategy for Artificial Intelligence in Science, deeply anchored in the EU’s long‑standing open‑science policy. Its policy framework emphasizes open data, reproducible research outputs, and ethical governance. The EU invests heavily in pan‑European public computing and dataset infrastructure for AI‑for‑Science, while adopting a more cautious stance toward the use of closed‑source commercial models in publicly funded research outputs.

3.4 Insights from Cross‑National Comparison

According to a 2023 OECD report on AI and scientific research, no universal optimal development path exists. National strategic differences reflect each region’s research‑and‑industry endowments: where enterprises possess abundant computing resources, private firms play a larger role in model development; where open‑science traditions prevail, public infrastructure and open sharing receive higher priority.

National strategies set out ambitious blueprints, yet policy documents do not guarantee real‑world implementation. Shifting focus from policy to individual disciplines reveals wide disparities in AI‑for‑Science maturity across subject areas.

4. Real‑World Implementation: Uneven Maturity of AI‑for‑Science across Disciplines

Observations of global projects show highly uneven practical adoption of AI‑for‑Science across research fields.

4.1 Disciplines with Routine Production‑Grade Adoption: Biomedicine and Materials Science

Biomedicine and materials science deliver the most tangible real‑world outcomes to date. These fields host large‑scale public experimental datasets, and their research problems readily translate into prediction and screening tasks. AI models are widely deployed for large‑scale prescreening of candidate molecules and materials, filtering out unpromising candidates and drastically cutting experimental trial‑and‑error costs. Many research groups have integrated model‑based prescreening into standard research workflows.

4.2 Primarily Supporting Acceleration: Climate Modelling

Climate modelling represents another representative use case. AI‑for‑Science models accelerate and reduce the complexity of conventional climate‑simulation frameworks. While preserving acceptable accuracy, they substantially cut computational overhead and support certain climate‑risk‑assessment workflows. Nevertheless, high‑fidelity simulations of complex extreme‑climate events still rely heavily on traditional physical‑simulation foundations; AI acts mainly as an accelerator rather than a full replacement for physical models.

4.3 Fields Still Limited to Proof‑of‑Concept Demonstrations in Basic Science

Many basic‑science domains remain constrained by shortages of high‑quality annotated research data. Here, AI‑for‑Science outputs mostly constitute proof‑of‑concept demonstrations. While small‑scale demos prove technical feasibility, such systems are not yet deeply embedded within day‑to‑day research pipelines.

4.4 Distinguishing Demonstration Benchmarks from Routine Research Production

These cross‑domain disparities call for a two‑tier understanding:

1. Demonstration‑level benchmark results: prove technical feasibility but remain distant from everyday research practice.
2. Routine research‑production‑grade applications: fully integrated into researchers’ regular workflows and consistently generating standard research outputs.

The two categories must not be conflated.

Note: The above summary of disciplinary implementation gaps draws from global public projects and reviews in the literature.

High‑profile demo results attract attention, yet routine production‑ready deployment is the true yardstick for paradigm maturity.

This uneven landscape emerges from interactions among multiple stakeholders. Governments, research institutes, corporations, and front‑line researchers each hold distinct priorities, jointly driving AI‑for‑Science forward while generating inherent industry tensions.

5. Perspectives from Diverse Stakeholders: Opportunities and Misaligned Incentives among Governments, Research Institutions, Enterprises, and Researchers

AI‑for‑Science advances through collective efforts of multiple stakeholders. Divergent core objectives and resource endowments create intrinsic industry tensions that shape today’s AI‑for‑Science ecosystem.

5.1 Government Research Agencies: Strategic Positioning, Public‑Resource Coordination, and Research Governance

Government research authorities view AI‑for‑Science as a critical frontier in global technological competition. Their core objectives include securing technological advantages, building public computing and dataset infrastructure, guiding resource allocation via policy, and advancing research governance that balances innovation speed, reproducibility, and talent development. That said, large‑scale computing and dataset infrastructure carry steep costs, creating persistent challenges for public‑resource allocation.

5.2 Universities and Public Research Institutes: Access to Scientific Questions and Experimental Infrastructure, yet Shortfalls in Engineering Capacity

Universities and public research institutes possess abundant research questions, experimental platforms, and expert researchers. Researchers hope AI‑powered research tools can reduce repetitive labour in large‑scale screening and simulation, freeing time for problem formulation and result interpretation. Significant practical barriers persist nonetheless: high barriers to acquiring high‑quality research datasets, and limited AI‑engineering capacity within many traditional research groups. Adopting AI‑for‑Science requires recruiting interdisciplinary staff, creating tangible bottlenecks.

5.3 Technology Enterprises: Strengths in Computing and Engineering, Misalignment between Commercial Goals and Open‑Science Expectations

Technology companies bring substantial computing resources and large‑model‑engineering expertise, making them key participants in AI‑for‑Science. Enterprises seek to validate model capabilities by solving real‑world scientific challenges and explore commercialization pathways such as technology licensing and research services. However, corporate commercial priorities naturally conflict with public‑science norms emphasizing openness and reproducibility. Some landmark results are produced using closed‑source models, making full experimental replication difficult for external researchers and introducing new challenges for research evaluation.

5.4 Front‑line Researchers: Hopes for Improved Efficiency alongside Caution toward Black‑Box Models

OECD reports document ambivalence among working researchers. On one hand, they welcome efficiency gains enabled by models. On the other hand, many remain cautious about outputs from black‑box systems due to concerns over reproducibility and insufficient mechanistic explanation. They refrain from treating model outputs as definitive scientific conclusions.

Misaligned stakeholder priorities do not equate to outright conflict. Industry‑university‑research collaboration represents today’s primary advancement model for AI‑for‑Science: research institutions contribute scientific questions and experimental‑validation infrastructure, while enterprises supply modelling expertise and computing resources for complementary cooperation. Still, misaligned incentives persist and constitute an essential context for understanding industry dynamics.

Transformations brought by model‑driven science extend beyond technology. They reshape research workflows, industry‑university‑research collaboration patterns, and expectations for researcher competencies.

6. Ripple Effects of Paradigm Change: Shifts in Research Workflows, Industry‑Academic Collaboration, and Required Researcher Competencies

Paradigm‑driven change entails more than introducing new tools. It reshapes end‑to‑end research workflows, transforms industry‑academic‑research collaboration models, and imposes new competency requirements on the next generation of researchers.

6.1 Structural Shifts in Research Workflows

In conventional research workflows, researchers formulate hypotheses, design experiments, run trials, and analyse experimental data. With model‑driven research in place, workflow components shift: researchers define scientific objectives, and models generate large‑scale candidate outputs and conduct virtual prescreening. Researchers then design real‑world experiments to validate model‑produced candidates. Models take on heavy‑duty trial‑and‑error screening work, allowing researchers to focus more on problem definition, experimental validation, and interpretation.

AI‑for‑Science redistributes research labour rather than eliminating human reasoning and critical judgement.

6.2 Industry‑Academic‑Research Collaboration Shifts from Late‑Stage Translation to Early‑Stage Deep Integration

Workflow transformations further reshape industry‑academic‑research partnerships. Historically, collaboration mostly occurred in late‑stage technology transfer: research institutions generated findings for enterprises to commercialize. In the AI‑for‑Science era, partnerships increasingly begin at the origin of research. Enterprises provide model‑and‑computing foundations, while research institutions contribute scientific questions and experimental validation. Both sides engage in deep collaboration from project inception onward. This places new demands on research‑project governance, intellectual‑property agreements, and data‑sharing frameworks.

6.3 New Competency Requirements for Research Talent

Paradigm shifts ultimately affect people. Traditional domain‑specialist researchers possess deep subject‑matter expertise, while AI specialists excel at algorithm design yet lack domain‑specific scientific knowledge. AI‑for‑Science creates pronounced demand for interdisciplinary talent: professionals who understand domain‑specific research questions, grasp model limitations, and can translate scientific problems into model‑ready tasks.

Universities worldwide are launching interdisciplinary curricula and training programs. Even so, talent development cycles are long, and shortages of qualified interdisciplinary personnel will continue to constrain AI‑for‑Science adoption across many research groups.

Given these shifts in workflows, collaboration models, and talent demands, what does the medium‑to‑long‑term future hold for AI‑for‑Science?

7. Looking Ahead: Two Technical Pathways and Medium‑to‑Long‑Term Outlook for AI‑for‑Science

Two major evolutionary pathways coexist within AI‑for‑Science today: general‑purpose scientific foundation models and domain‑specialized models. General scientific foundation models pursue cross‑disciplinary versatility with the vision of lowering barriers to adoption across fields, yet they incur high computational costs and suffer accuracy limitations on niche tasks. Domain‑specialized models are custom‑built for discrete subject‑area tasks, delivering superior performance within targeted use cases at the cost of generalizability; Returning or retraining is often required when switching research contexts. Evidence from global projects suggests the two approaches will likely coexist and complement each other rather than compete as mutually‑exclusive alternatives.

Drawing lessons from historical scientific‑paradigm evolution, paradigm transitions unfold incrementally. Short‑term bursts of disruptive fundamental‑science breakthroughs should not be expected. Technology accelerates experimental screening, virtual simulation, and similar steps. Nevertheless, defining meaningful scientific questions and interpreting scientific significance remain inherently human activities.

Models can speed up scientific discovery processes, but formulating truly valuable scientific questions stays the domain of human researchers.

Conclusion

Models deliver powerful new acceleration engines for scientific research — yet an engine itself is not equivalent to scientific discovery. The new cycle opened by AI‑for‑Science brings tremendous advances in tools and workflows. Even so, framing research questions, defining objectives, and interpreting scientific meaning continue to rely on human curiosity and critical reasoning. Paradigm evolution expands humanity’s boundaries of exploration rather than replacing human explorers.

References

[1] State Council of the People’s Republic of China. Opinions on Further Implementing the “AI +” Initiative (Guo Fa〔2025〕No.11)[Z]. 2025‑08.

[2] Jumper J, Evans R, Pritzel A, et al. Highly accurate protein structure prediction with AlphaFold[J]. Nature, 2021, 596(7873): 583‑589.

[3] National Academies of Sciences, Engineering, and Medicine. Foundation Models for Scientific Discovery and Innovation: Opportunities Across the Department of Energy and the Scientific Enterprise[R]. Washington, DC: The National Academies Press, 2025. DOI:10.17226/29212.

[4] OECD. Artificial Intelligence in Science: Challenges, Opportunities and the Future of Research[R]. Paris: OECD Publishing, 2023. DOI:10.1787/a8d820bd‑en.

[5] European Commission. A European Strategy for Artificial Intelligence in Science (COM(2025) 724 final)[R]. Brussels: European Commission, 2025‑10‑08.

[6] Hey T, Tansley S, Tolle K. The Fourth Paradigm: Data‑Intensive Scientific Discovery[M]. Redmond: Microsoft Research, 2009.