Review maps AI science loop and data power costs

A systematic review published in The Innovation describes a fundamental shift in how artificial intelligence is integrated into scientific discovery. The 60-page paper, authored by a team spanning 94 institutions, argues that AI has moved beyond being a task-specific tool for analyzing data. Instead, it now occupies a closed-loop workflow that links literature research, data analysis, hypothesis formulation, experimental execution, report generation, and autonomous evolution. This transition marks a departure from previous eras where AI, such as AlphaFold, solved specific prediction problems like mapping protein sequences to structures but left the broader scientific cycle to human researchers.

The review highlights a concrete example of this new capability through an autonomous laboratory known as A-Lab. In a specific test, A-Lab was tasked with synthesizing 58 target compounds. Over a period of 17 days, the system successfully produced 41 of these compounds. It achieved this by mining synthesis recipes from existing scientific literature and autonomously adjusting its course when a particular chemical route failed. The authors argue that this ability to execute and adapt within the experimental loop is the defining change in AI for science, rather than the development of a single, more powerful predictive model.

These closed-loop systems operate as integrated networks rather than linear pipelines. Language models provide domain knowledge and reasoning capabilities, while agents sequence the various steps of the research process. Embodied systems, such as robotic platforms, supply the physical capability to perform experiments. Self-correction mechanisms are built into these workflows through execution-and-evaluation pipelines and predict-measure-update cycles. These features are designed to catch errors before they propagate through the scientific process. Some research groups are currently testing whether AI systems can also participate in the peer review process itself.

The maturity of these AI-driven workflows varies significantly across the 11 fields surveyed in the review. The fields range from information science and mathematics to physics, chemistry, materials, medicine, life sciences, astronomy, geoscience, and social science. Materials science is identified as the field where the loop is closed furthest. In this domain, generative models propose new crystal structures, and robotic platforms are used to make and test these materials. In contrast, brain-computer interfaces offer a smaller but cleaner case study. A memristor-based decoder in this area gained about 20% accuracy over six-hour sessions by adapting alongside the user’s own neural signals.

To evaluate the validity of claims made by these AI systems, the authors propose a three-level test for evidence. The first level is offline benchmark performance, which measures how well a model performs on standardized tests. The second level is domain-grounded agreement with accepted science, ensuring the AI’s outputs align with known physical and biological laws. The third and most rigorous level is prospective or experimental validation in a real setting. The review emphasizes that only this third level of evidence settles whether a result constitutes genuine science rather than just a high score on a benchmark.

Despite the progress, the review identifies three major constraints that bind the field. First, standardized, high-quality data remain scarce and long-tailed, limiting the ability of models to generalize. Second, deep neural networks often trade interpretability for accuracy. A model may predict a molecule’s activity with high precision without explaining the underlying chemical reasons for that activity. Third, generative models can output structures that appear plausible but violate fundamental conservation laws or valence rules. The authors refer to this phenomenon as physical hallucination. A fourth risk is human-centric: the fluency of AI output can be mistaken for genuine understanding, leading to over-reliance on the system.

The scale of AI for science carries a significant environmental and energy cost. Data centers drew about 415 terawatt-hours of electricity in 2024 and are on course to roughly double that usage by 2030. This projected increase is approximately equal to Japan’s current annual electricity use. More than half of the power used by US data centers is still derived from fossil fuels. The review also notes the rapid policy response to this expansion, citing the US Genesis Mission, Britain’s and the EU’s AI-for-science strategies, and China’s AI+ push, all launched within a single year. The authors argue that the advantage in this field will go not to the entity with the largest budget, but to whoever best integrates data, compute resources, and human expertise.

The review serves as an update to a map published by the same team in 2021 in the same journal. It provides a comprehensive overview of the progress, challenges, and perspectives for AI in science. The authors declare no competing interests. The study underscores that while AI is becoming a foundation for scientific discovery, significant barriers remain in data quality, interpretability, physical consistency, and the establishment of trusted human-AI research partnerships. The field is moving toward a future where AI not only assists but actively drives the scientific method, necessitating new standards for validation and trust.

Subject of Research: Systematic review of AI for science closed-loop workflows and challenges

Article Title: AI now writes the hypothesis, runs the experiment and grades itself. A 60-page review asks how much of that science we can trust

Article References: AI now writes the hypothesis, runs the experiment and grades itself. A 60-page review asks how much of that science we can trust. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: AI for science, closed-loop systems, autonomous laboratory, systematic review, The Innovation, data centers, energy consumption, scientific discovery, generative models, peer review, writes, hypothesis