Artificial Intelligence Is Moving From Physics Assistant to Experimental Partner
For centuries, discovering a new law of physics meant some combination of human intuition, mathematical insight, painstaking experimentation and, occasionally, extraordinary luck.
That formula is beginning to change.
Artificial intelligence systems are now being trained not merely to analyze scientific papers or calculate equations, but to search enormous spaces of possible materials, generate hypotheses, operate laboratory instruments and test predictions against experimental evidence.
The most intriguing developments are emerging in quantum materials and superconductivity, where researchers are using machine learning to identify promising materials that could otherwise take years to find. At the same time, new AI systems are being designed to reconstruct physical laws from raw experimental data rather than simply being told which equations to use.
The technology is still experimental. AI has not suddenly replaced physicists, nor has it independently discovered a revolutionary new law of nature.
But the direction of research is significant: scientists are beginning to build systems capable of participating in parts of the scientific discovery loop themselves.
The Search for Materials Humans Cannot Easily Find
Superconductors are among the most fascinating materials in modern physics.
Below a critical temperature, a superconductor can carry electrical current with essentially zero electrical resistance. Some superconducting systems also exhibit quantum effects that could be important for quantum computing, advanced sensors, medical technologies and future energy infrastructure.
The problem is finding useful superconductors.
The number of possible combinations of elements, crystal structures and electronic configurations is enormous. Traditional materials research can therefore become a painstaking process of theoretical calculations, candidate selection, laboratory synthesis and testing.
AI offers a different strategy.
Instead of asking researchers to examine candidates one by one, machine-learning systems can search enormous databases and prioritize materials that appear to possess particular physical properties.
That approach is already producing experimentally verified results.
AI Has Already Helped Find New Superconductors
A 2025 research project involving scientists from the University of Florida and Oak Ridge National Laboratory demonstrated how machine learning could become part of an end-to-end superconductor discovery pipeline.
The researchers developed a model called BEE-NET to predict properties associated with superconductivity and integrated it with computational methods, elemental-substitution strategies and machine-learned interatomic potentials.
The system began with approximately 1.3 million candidate structures and narrowed the search to 741 compounds that were stable and had density-functional-theory-confirmed superconducting critical temperatures above 5 kelvin.
Most importantly, researchers reported successfully synthesizing and experimentally confirming two previously unreported superconductors from the AI-assisted search.
That distinction matters.
The AI did not simply produce a list of imaginary materials.
Its predictions were used to guide physical experiments, and two candidates survived the journey from computer model to laboratory sample.
This is the emerging model of AI-driven physics: algorithms reduce the search space, physics calculations test plausibility, and experiments determine whether nature agrees.
The Next Target: Topological Superconductors
The challenge becomes even more complicated when researchers search for topological superconductors.
Topological materials possess unusual electronic properties arising from the mathematical structure—or topology—of their quantum states. When superconductivity is combined with topological behavior, researchers become interested in exotic states that could potentially support Majorana modes.
These systems are particularly important because Majorana-related states have long been investigated as possible ingredients for more robust approaches to quantum information processing.
But identifying a genuine topological superconductor is exceptionally difficult.
Researchers must consider crystal symmetry, electronic structure, superconducting behavior and topology simultaneously.
A presentation at the 2026 American Physical Society Global Physics Summit described an AI-based framework designed specifically for this challenge. The system, called a “TSC Materials Identifier,” uses ab initio band information to identify candidate topological superconductors and determine the expected topological character of their superconducting phases. It can also classify possible boundary states, including edge, surface, hinge and corner behavior.
That could dramatically change how researchers approach the field.
Rather than starting with a handful of materials that happen to look promising, scientists can potentially ask an AI system to search a much larger materials landscape for specific quantum characteristics.
From Prediction to Autonomous Science
The bigger development, however, goes beyond machine-learning prediction.
Researchers are increasingly connecting AI systems directly to laboratory instruments.
This creates what scientists call a self-driving laboratory.
The basic concept is straightforward.
An AI system proposes an experiment.
Automated equipment performs it.
Sensors generate data.
The AI analyzes the results.
It then decides what experiment should happen next.
The process repeats.
Instead of an algorithm producing one prediction for a human scientist, the AI becomes part of a closed experimental feedback loop.
A 2025 Nature Communications study demonstrated both the promise and limitations of this approach using an AI laboratory assistant called AILA. Developed by researchers including scientists at IIT Delhi, AILA was designed to operate an atomic force microscope, control experimental parameters, analyze results and adapt its workflow.
The researchers reported that a task that previously could take roughly a day to optimize could be reduced to around seven to 10 minutes in their experimental setup.
That is a remarkable acceleration.
But the same study delivered an important warning.
The AI systems did not automatically become expert scientists simply because they could operate scientific equipment.
The researchers found that models capable of answering scientific questions could still struggle when confronted with unpredictable laboratory situations. Multi-agent systems performed better in some evaluations, but the study also identified instruction sensitivity and safety concerns.
In other words, autonomous science is advancing—but it is not solved.
Can AI Actually Discover a New Law of Physics?
This is where the story becomes more controversial.
AI systems have become remarkably good at identifying patterns. But identifying a pattern is not necessarily the same thing as discovering a physical law.
A scientific law must generally survive repeated testing, describe a broad class of phenomena and connect observations through a coherent physical framework.
Researchers are now experimenting with AI systems designed to take that next step.
One prominent example is AI-Newton, developed by a research team led by scientists at Peking University.
The system was designed to derive physical laws from experimental data without being explicitly given the relevant physical equations. In proof-of-concept experiments involving Newtonian mechanics, it successfully reconstructed principles including Newton’s second law, conservation of energy and universal gravitation.
That is significant—but it needs to be described accurately.
AI-Newton did not discover a previously unknown fundamental law of physics.
It rediscovered laws that humanity already knows.
The achievement is therefore better understood as a demonstration that AI can reconstruct general physical relationships from data and organize them into concepts and laws.
The next question is whether similar systems can move beyond rediscovery.
The Difference Between Rediscovering Physics and Finding New Physics
This distinction could determine whether AI becomes a genuine scientific revolutionary.
If an AI system is trained on existing scientific knowledge, it can become extremely effective at finding relationships that humans have already documented.
But genuinely new physics requires something harder.
The system must identify an unexpected relationship that is not simply a rearrangement of information already contained in its training data.
It then has to formulate a testable hypothesis.
Researchers must design an experiment capable of distinguishing that hypothesis from existing theories.
And nature must ultimately confirm the prediction.
That is a much higher standard.
A 2026 perspective on AI and physics has highlighted precisely this challenge: when AI predictions move beyond familiar examples, researchers must distinguish potentially meaningful physical discoveries from artifacts such as overfitting.
The danger is that an AI model can produce a mathematically impressive relationship that works on the data used to find it but fails elsewhere.
Physics does not reward clever curve-fitting.
Nature gets the final vote.
Why Autonomous Agents Could Change the Scientific Method
The real revolution may not be a machine suddenly inventing Einstein-level theories.
It could be something more practical.
AI agents can search more possibilities than individual researchers.
They can analyze enormous datasets.
They can repeat computational experiments continuously.
They can operate laboratory equipment.
They can compare competing hypotheses.
And they can potentially coordinate several specialized AI systems, each responsible for a different part of the scientific workflow.
Recent research is pushing toward precisely this model. A 2026 preprint describing a multi-agent system called AHOIS explored a closed-loop approach in which AI agents generated hypotheses, challenged them, proposed experiments and used evidence to revise their explanations in a physical optical system. The researchers described the work as a step toward evidence-grounded autonomous scientific discovery.
Such systems are still early-stage research.
But they reveal where the field is heading.
The future laboratory may not simply contain scientists using AI.
It may contain scientists supervising networks of specialized AI researchers and automated instruments.
What This Could Mean for Quantum Technology
The implications for quantum technology are particularly significant.
Discovering better superconductors could improve the materials used in quantum devices.
Finding topological phases could potentially reveal new ways of manipulating quantum information.
Understanding unusual quantum states could expose physical mechanisms that conventional theories have difficulty describing.
And faster materials discovery could reduce the time between theoretical prediction and laboratory validation.
The potential applications extend beyond quantum computing.
Advanced superconductors could influence sensors, medical imaging, energy systems and transportation.
Yet none of these applications should be treated as guaranteed outcomes of current AI research.
The distance between a promising computational prediction and a commercially useful material can be enormous.
The Human Scientist Is Not Disappearing
Despite the rapid progress, humans remain central.
Scientists determine which questions matter.
They establish experimental constraints.
They evaluate whether an AI-generated hypothesis is physically meaningful.
They design validation experiments.
And, critically, they decide whether an apparent anomaly represents new physics or simply an error.
The AILA research provides a useful reminder. AI agents can perform impressive laboratory tasks, but they can also behave unpredictably and remain sensitive to how instructions are structured.
Autonomous science therefore requires safeguards, reproducibility standards and independent verification.
A discovery that cannot be reproduced is not a new law of nature.
The Beginning of an AI-Driven Physics Era
The most important development may be happening quietly inside laboratories rather than in headline-grabbing demonstrations.
AI is moving from a tool that helps scientists calculate faster toward a system that can participate in the scientific method.
It can search for superconductors.
It can identify topological phases.
It can operate instruments.
It can generate hypotheses.
It can reconstruct known physical laws from raw data.
And increasingly, researchers are testing whether AI can design experiments that challenge its own explanations.
The extraordinary claim that AI will soon discover entirely new laws of physics remains unproven.
But something more concrete has already happened: the boundary between computation, experimentation and scientific reasoning is becoming increasingly blurred.
The next major physics breakthrough may therefore begin not with a physicist writing an equation on a blackboard, but with an AI agent noticing a pattern buried inside millions of possibilities—and asking the one question that no human researcher thought to ask.
