SUNDAY, 7 DECEMBER 2025
Imagine the year 2075. As we look back on the 2020s, this decade would stand out as a pivotal moment in the history of science defined by the rise of machine learning and artificial intelligence (AI). The 2024 Nobel Prize in physics and chemistry and the exploding popularity of large language models are the smoking gun of this milestone. Analogous to the Scientific Revolution of the 16th to 18th centuries—marked by paradigm-shifting breakthroughs from Copernicus, Newton, and Galileo—this new era is increasingly being shaped by AI. The Scientific Revolution fundamentally transformed people’s attitude to the natural world by overturning long held religious beliefs and teachings, through the development of the scientific method based on observation and reasoning. It gave us a new framework to understand the world which paved the road for modern science. Drawing parallels to the present, AI has given us a new tool to understand the world through finding patterns in massive and unstructured datasets, reformulating existing problems and generating predictions of new content. Of course, AI is not a novel topic in the scientific community and has been adopted in various form for decades. With increased computing power and larger datasets, it will be an exciting time going forward to witness what sort of new scientific discoveries that could be enabled by AI.AI ENABLES NEW IDEAS AND APPROACHES TO SCIENTIFIC PROBLEMS.
Often, challenging theoretical problems in physics involve solving complex systems of partial differential equations (PDE), searching through vast parameter space or performing long and tedious calculations.
String theory, a theoretical framework attempting to unify quantum mechanics and general relativity – our leading theories describing interactions at the smallest and largest scales in the universe, respectively – hosts many of those challenges. A key area of research in string theory involves investigating the shapes and properties of extra dimensions to understand how they affect the physical properties of particles and forces. Starting with 11 dimensions, there are an enormous number of ways to fold up the higher dimensions to give rise to the 4-dimensional universe we see. This process involves solving numerous complex and nonlinear PDEs. Development of powerful machine learning techniques has inspired researchers to reformulate the PDE solutions as optimisation problems obeying certain constraints. One can see the power of machine learning in providing a general framework for solving seemingly unconventional problems. This has allowed researchers to perform calculations at a much faster rate than before, accelerating progress in the field.
Moreover, AI is streamlining the synthesis of information. An important element of science research involves having a broad understanding of relevant research done by others, and many a time, this involves scouring and digesting numerous academic papers. In addition, a key source of creativity in research comes from drawing parallels from seemingly distant fields of research. In this spirit, large language models (LLMs) are increasingly being used to accelerate or automate such laborious text-based tasks like conducting literature reviews.
To name a few, tools such as Semantic Scholar and Elicit can analyse thousands of academic papers in minutes, identifying key insights and cross-disciplinary connections. Instead of drowning in academic readings, researchers can now, in theory, focus their mental prowess on asking the right questions and coming up with creative solutions to complex problems. Where LLMs lag behind their human counterparts, however, is in their inability to conduct critical thought, to assess bias and manipulation, to examine images and figures. As such, outputs from these models, although they may be presented with great confidence, are not yet completely accurate or trustworthy which is a key concern for those wishing to rely solely on them for such tasks.
AI PERFORMS ADVANCE COMPUTATION, SIMULATION AND DATA ANALYSIS
AI’s computational ability has already catalysed major breakthroughs. One of the poster children of its success is AlphaFold, an AI system developed by Google DeepMind to address the protein-folding problem. Predicting how a sequence of amino acids folds into a three-dimensional structure has been a long-standing challenge for biologists. Previously, experimental determination of a single protein structure could take years, involving painstaking efforts in the laboratory. AlphaFold has revolutionised this field by using deep learning to predict protein structures up to atomic-level accuracy. It does so by leveraging on the availability of large amounts of experimentally determined protein structures in the Protein Data Bank. More importantly, Anfinsen’s thermodynamic hypothesis, which states that protein folding minimises free energy regardless of the path to get there, allows us to cast it as an ideal problem for deep learning with clear objective function.
Data-intensive fields like astronomy and particle physics are experiencing AI-driven transformations. For example, the James Webb Telescope generates 235 gigabytes of data every day and of course, the Large Hadron Collider, when running, generates one petabyte of collision data per second (equivalent to 500 billion pages of text). Researchers in these fields apply AI to sift through the massive amounts of information, leveraging on the ability of machine learning models to identify patterns and anomalies among datasets. AI-based workflows not only accelerate discovery but also free researchers to focus on interpreting findings and developing new theories.
Looking ahead, AI welcomes a transformative era in scientific research, offering substantial opportunities across many scientific fields. While AI can be a powerful tool, new issues arise regarding reproducibility, equitable adoption, and the synergy between human expertise and AI systems. Nevertheless, this AI-driven scientific revolution can redefine discovery, and offer solutions to some of the world’s most pressing challenges. At the end of day, remember that human creativity and judgment is what puts us in the driver’s seat of this revolution.
AI AS INDEPENDENT SCIENTIFIC AGENTS
Automated experimentation is a developing prospect in materials research to perform iterative experimental processes using AI and robotics in a closed experimental loop. By defining an objective and specifying a clear workflow, autonomous systems can perform routine and time-consuming experiments. This frees up researchers to work on other tasks and speed up developments, which could mean technologically relevant materials can be made commercially available in a shorter time. For instance, there has been research to automate and optimise the growth of carbon nanotubes. Such AI agents could also analyse the results and propose new experiments, aiding researchers to discover new research directions. As with many areas of AI development, this is an ongoing, divisive issue, with great excitement from some camps and great, and not unfounded, concern from others.
CHALLENGES OF AI IN SCIENCE RESEARCH
While AI hold great promises, it brings with it numerous challenges as well. The most widely recognised issue with AI is its black-box and non-transparent nature, a consequence of its complex architecture and opaque internal decision-making process. This could lead to problems with verification and reproducibility of results, which are the basis of trustworthy scientific research. Worse still, the limited transparency of commercial models that power AI-based research is something to worry about too. In addition, overreliance on the use of AI might foster a sense of complacency and damage creativity. Furthermore, as high-quality data becomes a key driver of new scientific discoveries, equitable data access becomes ever more important so that scientific research does not concentrate among a few selected research groups. Lastly, placing a superficial premium on AI in terms of getting publications might damage the quality of science research. People should not forgo “good science”, even if it means using conventional methods, for the sake of being “good at AI”.
Shikang Ni is an MPhil student for Scientific Computing at Wolfson College, with a keen interest in fusion energy and AI technology. He is excited about emerging technologies and hopes to engage people.
