Why science will be AI's killer app
Artificial intelligence has begun to automate the process of discovery with significant implications for science, scientists and society
The great surprise with artificial intelligence is its ability to perform seemingly complex tasks, sometimes better than humans. There are numerous visions of an AI future in which machines manage our diaries, our holidays, our relationships, our jobs and more.
But while large language models (LLMs) show potential in all these areas, their true worth lies in the creation of new knowledge about the world. This is currently the role that science plays in society. So an important question is whether AI can move beyond regurgitating what is known to discovering what is not.
A new generation of AI systems is beginning to answer that question. These systems are automating the process of knowledge discovery by performing the same tasks previously carried out by scientists. This raises the prospect that AI’s best use will not be in becoming intelligent assistants or relentless workers or sympathetic friends or anything as mundane as this. Increasingly, the message emerging from AI companies is that AI’s killer app is science itself.
Relentless innovation
Laboratory science is largely the methodical, relentless pursuit of evidence in controlled conditions. That is a difficult task, not because of a single, insurmountable obstacle, but due to a thousand tiny, complex ones.
Now Shuxiang Cao at the University of Oxford and colleagues have developed an AI system that can automate the complex workflow of a laboratory.
That’s no easy task. Laboratory procedures are often updated and not usually publicly available, making it difficult to train most LLMs on them. The specialised nature of the work means the volume of data is insufficient to fine-tune the output of LLMs. And much laboratory know how is not written down any but exists purely in the minds of the researchers who gather it over months or years of experience.
“Despite challenges, pioneering efforts to develop automated LLM-based agents to carry out experiments are already underway,” say Cao and co. Their approach takes this work further.
The team have developed a set of knowledge-based agents or “k-agents” that work together on different aspects of laboratory work to perform an experiment from start to finish. They put the agents through their paces in a quantum computing laboratory to measure and calibrate a superconducting qubit, a process that is usually a significant bottleneck in the field of quantum information processing.
This calibration—finely tuning the dozens of parameters for each qubit—is time-consuming, repetitive and labour-intensive. As such, it relies heavily on the intuition and experience of human scientists and involves a dizzying array of knowledge types. A scientist needs to know how to translate a high-level goal into a sequence of concrete experimental steps, how to write the code to control the apparatus for each step, how to interpret the often-graphical results to see if a step worked and decide what to do next based on that outcome.
The k-agents tackle this with a variety of expertise in a specific domain. For instance, Code Translation Agents convert natural language instructions into the precise Python code needed to operate the machinery. An example would be turning the phrase "Run a Rabi experiment" into code that executes a series of experimental steps that zap a qubit with a pulse of microwave radiation at its resonant frequency, causing it to oscillate between its ground and excited states.
Then there are Procedure Translation Agents, which understand more complex, multi-step workflows and can break them down into a logical sequence and write code to apply them.
Perhaps most critical are the Inspection Agents. After an experiment is run, these agents analyse graphical plots of the experimental results to determine if it was a success.
Orchestrating this team is the Execution Agent that decides what to do next: move to the next stage or retry the last stage with adjusted parameters.
This structure is a key breakthrough, say Cao and co, because as it allows the system to tackle long and complex tasks without being overwhelmed.
Quantum processor
To prove their system’s mettle, the researchers turned it loose on a superconducting quantum processor, one of the leading platforms for building a universal quantum computer.
First, the system had to fully calibrate single qubit gates. The AI correctly decomposed this into the necessary sequence of calibration steps known as Ramsey, Rabi, and DRAG experiments.
At one point, an inspection agent reviewed the data from a Ramsey experiment to determine the coherence time of a qubit and concluded that the plot didn't show enough oscillations to make a good estimate.
Acting on this report, the execution agent autonomously decided to repeat the experiment with a longer runtime, which then produced a successful result. In the end, the system reported a fully calibrated, high-fidelity qubit.
Next, the team tasked the AI with a more challenging discovery problem: finding the optimal parameters for a two-qubit entangling gate, a crucial component for any quantum computer. This task typically involves a painstaking search through a vast parameter space of driving frequencies and amplitudes, usually guided by a scientist's empirical knowledge.
Instead the system autonomously proposed parameter sets, ran experiments and used its inspection agents to evaluate the complex graphical results. After 100 experiments over three hours, the system identified an optimal set of parameters to reliably entangle two qubits.
Finally, to demonstrate the system could use the components it had just built, the researchers instructed it using natural language to prepare a three-qubit entangled state called a Greenberger-Horne-Zeilinger state, a foundational quantum information task. The agent successfully translated the request into code, ran the experiment using the gates it had just calibrated and reported a state fidelity of over 83 per cent.
Of course, the system isn't perfect. It relies on clean, well-documented codebases and procedures which are time consuming to produce. “Such structured knowledge may not always exist, and transforming existing knowledge could take considerable human effort,” say Cao and co. But they add that “the setup effort is comparable to preparing laboratory manuals or protocols for new PhD students or junior staff.”
In comparison to humans, the LLMs take longer to carry out tasks like evaluating data plots but are significantly faster at producing code. “Overall efficiency is comparable,” say the team. But they expect improvements as AI models and infrastructure improve.
"These achievements suggest that our system is a valuable tool for research groups working with superconducting quantum processors, with the potential for broader applications in research automation across other fields," say Cao and co, adding that the framework is general enough to be adapted to other automatable laboratory settings in physics, chemistry, biology and materials science.
The team’s vision is of a "self-driving laboratory," where AI systems can execute experiments and then hypothesise, plan and learn from the results in a closed loop of discovery. That’s a fundamental breakthrough which will allow AI to actively and intelligently participate in the scientific process.
INSIGHT
This work paves the way for a very significant change in the nature of science and the democratisation of laboratory automation across many fields. The k-agents framework provides a blueprint for moving beyond rigid, hand-coded automation scripts, which require significant programming expertise, to a more flexible and intuitive system.
This will lower the barrier to entry for high-throughput experimentation, enabling more research groups to tackle complex, long-duration tasks that were previously too labour-intensive.
The shift will also redefine the role of the scientist in the laboratory. Instead of being a hands-on operator constantly monitoring experiments and adjusting parameters, the scientist will evolve into a high-level supervisor and strategist. Their primary tasks will be to define the overarching scientific goals, curate the high-quality documentation and code that the agents learn from and intervene when the AI encounters novel situations or produces unexpected results.
This frees up researchers from time-consuming and repetitive work, allowing them to focus on the more creative aspects of science: forming hypotheses, interpreting complex data and designing the next wave of inquiry.
It raises the prospect of a new kind of discovery, analogous to the emerging technique of “vibe coding” enabled by platforms like Replit and Cursor. These generate sophisticated apps and programs based on natural language prompts—DIY coding for the masses.
But the scientific equivalent could be even more transformational. “Vibe science” would allow researchers to immediately test ideas and hypotheses developed in brainstorming sessions, by intuition or guess work.
All this suggests the pace of scientific discovery is set to accelerate as self-driven labs close the loop between hypothesis, experimentation and analysis. And do it round the clock.
In the near term, the focus will be on improving the initial training and knowledge transfer to k-agents. Beyond that, a key goal will be to develop AI capable of structuring this knowledge automatically from existing, messy project files with just some human guidance. The result will be a virtuous cycle where the AI not only performs the science but also organises the knowledge required to do so more efficiently.
And what of scientists themselves? Only a few years ago, companies were crying out for coders with the skills to create the next generation of AI-enabled programs and apps. But hiring rates have dropped dramatically recently as LLMs have increasingly taken on this work themselves. Coding skills are no longer the guarantee of well-paid work they once were.
Will the same be true for scientists? Will science really be AI’s killer app? On this evidence, it’s hard to rule out.
Ref: Agents for self-driving laboratories applied to quantum computing : https://arxiv.org/abs/2412.07978