91´«Ã½

Image caption: Illustration of how a particle's shape influences the flow patterns around it.

For seven years,ÌýAnkur Gupta, assistant professor of chemical and biological engineering at 91´«Ã½, had been pursuing a theory about how small particles move when driven by electric fields. Century-old research suggests that particle size and shape do not change their speed, but Gupta wanted to go deeper. He sought to understand how size and shape affect the motion of smaller particles when driven by electric fields, and with the help of Claude, an AI assistant, and his team’s rigorous verification, the researchers solved the problem in just five weeks.

In the following Q&A, Gupta discusses the research, how AI contributed to the discovery and the importance of human oversight.

Claude helped you solve a problem in five weeks that had remained unsolved for decades. What problem were you trying to solve?

There’s an intense interest in fabricating micron- and submicron-sized particles with a wide range of shapes. These tiny engineered particles are increasingly used in medical diagnostics, targeted drug delivery and tiny self-propelled devices, where precise control of their movement is essential.

I was trying to develop a theory that predicts how these small particles move when subjected to an electric field and how their size and shape affect their motion.

What was still unknown about how small particles move?

Imagine kicking a ball. How fast it travels depends on how hard you kick it, as well as the ball’s size and shape. Something similar happens at the microscopic scale in a process called electrophoresis, where tiny charged particles move through an electrolyte solution when an electric field is applied. The electric field acts like the kick: the stronger the electric field, the faster the particle moves.Ìý

One might expect that, like a ball, larger or differently shaped particles would move at different speeds during electrophoresis. The Polish physicist Marian Smoluchowski showed more than a century ago that the speed of particles larger than a few micrometers is independent of size and shape. The remaining challenge was to understand what happens when particles become small enough for their size and shape to affect how they move.

What did you discover about how particle shape affects motion?

We found that the most important aspect of a particle’s shape is the stretch or compression, the change from a perfect sphere into a slightly elongated or flattened shape. Other deformations, such as waviness, have little effect on how the particle moves. Being able to predict how shape controls this motion is important because engineered micro- and nanoparticles are increasingly used in medical diagnostics, targeted drug delivery and self-propelled "micro/nano swimmers," where controlling their movement is essential.

What potential impact could your findings haveÌý on applications such as medical diagnostics and targeted drug delivery

Our findings show that a nanoparticle's shape affects how it moves in an electric field. This insight could help researchers design nanoparticles that travel at the right speed for applications like medical diagnostics and targeted drug delivery.

Why did you bring AI into the project?

For more than seven years, I had been pursuing a theory beyond the Smoluchowski result. My research group had explored several different approaches to the problem over a period of a year and a half, but we couldn’t find a solution.Ìý

There’s been a lot of buzz around AI, so we decided to give it a shot. We set up the problem, but we were curious whether AI could handle the long, detail-intensive algebra required to solve it.

What kinds of tasks was Claude especially good at?

It was especially good at handling repetitive, time-consuming tasks, such as carrying out lengthy calculations, writing computer code and creating publication-quality figures. Even then, we had to carefully review its work to make sure everything was accurate.

What were the dangers of bringing AI into research?

AI can make mistakes, and relying on it too much can spread those mistakes throughout a project. For example, AI may confidently execute a well-documented procedure in the wrong situation and provide explanations that justify an incorrect answer. Researchers need to be skeptical of anything AI does and verify AI-generated results independently.

What surprised you most about the mistakes Claude made? Were they difficult to spot?

AI made the kinds of mistakes we expected, such as misinterpreting research papers or relying on unrealistic assumptions. As the work became more refined, the mistakes became much harder to detect. Claude sometimes made subtle mathematical errors that appeared correct, or adjusted its reasoning to match expected results, creating answers that looked self-consistent but were wrong. In some cases, the results, including graphs, appeared valid until we carefully checked every step.

Also, when preparing the manuscript and an accompanying blog, we asked Claude to help draft the "mistakes" section of our blog post, and it fabricated three plausible-sounding errors that never occurred.Ìý

What does a scientist need to do to ensure AI-generated results are reliable?

Scientists need to be deeply involved in the process. They must carefully check AI-generated work against primary sources, their own calculations and their understanding of how the science should behave. No single check was enough in our case. Rather than replacing the scientist, AI changes the role: researchers can use it to help build solutions to more ambitious problems, but they must apply their expertise to verify, refine and correct the results.

Though you used AI, how much of your work still depended on your own expertise?

Ìý Ìý Ìý Ìý Ìý Ìý Ìý Ìý Ìý Ìý Ìý Ìý Ìý ÌýAssistant Professor Ankur Gupta

Expertise becomes even more valuable when using AI. While some tasks can be automated, other steps such as posing the problem correctly, choosing the right geometry and mathematical approach and interpreting what the results meant required our input. Claude's role was to accelerate the more time-consuming parts of the process, including calculations, coding and creating figures, but the scientific direction and decision-making remained ours.

Has this experience changed how you approach research?

It's made us much more deliberate about how we verify information. We now treat every AI-generated result the way we'd treat a result from a source we don't yet fully trust. It’s not something we can rely on until it has been checked against trusted sources or independent calculations.

What advice would you give graduate students about using AI responsibly?

Use AI to learn new techniques, but don't let it replace your own thinking. AI is excellent at combining existing methods, but much less capable of understanding a problem from a physical standpoint. That understanding must come from you, otherwise you will eventually lose the ability to evaluate AI's work. Think of it like learning chess. A player does not become a strong player by relying on a computer program that analyzes chess positions and suggests the best moves; you have to play thousands of games to develop the intuition needed to make decisions. Otherwise, you are not playing chess; you are simply following a computer's output.