At CERN's Large Hadron Collider, machine learning algorithms now decide which particle collisions to keep and which to discard. In materials science, AI-driven robotic labs have synthesized 41 novel compounds in 17 days. As artificial intelligence becomes embedded in the day-to-day work of physicists and materials scientists, a pressing question emerges: can science produce discoveries that researchers do not fully understand?
AI in particle physics: from Higgs to anomaly detection
Machine learning has been part of LHC research since the early 2010s, when physicists used it to separate Higgs boson signals from background noise. Those supervised algorithms were trained to recognize known signatures. Achieving the same sensitivity without ML would have required between 15% and 125% more data, with some measurements needing more than double the data.
Today, the scope is far broader. Neural networks compress enormous datasets, accelerate simulations, and flag unusual events among billions of collisions. Caterina Doglioni, an experimental particle physicist at the University of Manchester and member of the ATLAS collaboration, said AI "is pretty much everywhere" in high-energy physics. "It's critical at the LHC where far more data is created than can realistically be stored."
Self-supervised learning methods now preserve key features while dramatically reducing storage, reconstructing original data from compressed summaries. More recently, anomaly-detection techniques have moved from theory to real searches. Instead of hunting for a specific predicted signal, algorithms learn what ordinary collisions look like and flag deviations. In 2024, ATLAS reported results from an unsupervised anomaly search using Run 2 data. No deviation from Standard Model expectations appeared, but the study proved such methods work on real collider data.
The black box and reproducibility
The same pattern-matching power that makes these systems effective also creates unease. Modern ML models can uncover relationships that researchers struggle to explain. David Sutherland, a theoretical physicist at the University of Glasgow, draws a distinction: "I would certainly say reproducibility [is more important] than interpretability."
Particle physics has built extensive safeguards around that process. Doglioni pointed to large review committees that check "basically everything that goes on," providing a layer of protection against over-reliance on opaque systems. Yet as AI becomes more deeply integrated, reproducing results may require access not only to data and code but also to model architectures, training conditions and computational environments.
Christoph Weniger, a physicist at the University of Amsterdam, believes researchers will shift into "a supervisory role in steering these different agents," fundamentally changing how scientists interact with data. That shift could alter how scientific value is measured, placing more weight on the originality of questions rather than the volume of output.
AI in materials science: closing the loop
Particle physics is not alone. At Argonne National Laboratory, researchers built an autonomous system where AI proposes candidate materials, robotic labs synthesize and test them, and results feed back to guide the next experiment. In 2023, the A-Lab platform conducted 355 experiments in 17 days and synthesized 41 novel materials (Nature 624 86). This closed-loop approach shows AI not just analyzing data after the fact but deciding which experiments to run.
These developments are part of a broader trend in AI for Science & Research, where automated methods accelerate discovery. The same tools that make research faster also make scientific text easier to produce. arXiv recently announced one-year bans for authors who upload papers with obvious signs of unverified AI content, such as hallucinated references or leftover chatbot instructions. The move highlights growing concern about the quality of AI-assisted writing.
Training the next generation
If AI automates the routine tasks that once trained young researchers, the path to expertise becomes less clear. Wrishik Naskar, a particle theory postdoc at DESY, observed a shift in habits: "When I get stuck, I go to my supervisor. [But] many people, the first thing they ask is ChatGPT." His advice: "Focus on the learning. Results will follow if you have skills."
For researchers navigating this shift, structured resources such as an AI Learning Path for Research Scientists can help build the skills to work alongside these systems. Admir Greljo, a particle theorist at the University of Basel, sees collaboration rather than replacement. "It's not that they will compete against AI. Entering into theoretical physics is still going to be very interesting and very important. AI is not competition, it's a new tool."
Nichol Furey, a mathematical physicist at Humboldt University of Berlin, noted a different tension. "Authors in maths and physics often write in ways that are overly opaque," she said. "AI systems write with their audience in mind, whereas human authors often don't." The same tools that improve clarity also make it trivially easy to flood the literature with low-quality text.
Towards an AI scientist?
Sutherland can imagine a future where AI systems perform the entire research cycle, from hypothesis to paper, with a human "steering it" and acting as a sanity check. Greljo sees no fundamental barrier to fully autonomous scientific systems, arguing that reproducibility allows humans to verify whether the AI reached the right answer.
Science has always involved incomplete models and uncertainty. AI may help uncover truths faster than we can fully understand them, but that need not signal the end of inquiry. The problem with Douglas Adams' fictional supercomputer Deep Thought was never that it gave the wrong answer. It was that nobody knew what question they had asked.
Why this matters for science and research professionals
AI is not a replacement for the scientific method; it is a tool that shifts where human judgement is needed most. Reproducibility, careful validation, and the ability to ask precise questions remain central. For working researchers, the practical challenge is to develop enough fluency with AI systems to steer them effectively while preserving the deep problem-solving skills that come only from hands-on struggle. The scientists who thrive will be those who treat AI as a collaborator, not an oracle, and who keep their focus on deciding which questions are worth asking in the first place.
Your membership also unlocks:
