How accurately can you measure an object when all you have is a blurred image? Scientists at TU Wien, in collaboration with teams from the University of Glasgow and the University of Grenoble, have pushed the boundaries of what is physically possible—using artificial intelligence.

Figure 1. Edge of What Physics Allows.

The Fundamental Challenge of Blur

No imaging system, no matter how advanced, can completely eliminate blur. For more than 150 years, researchers have known that microscopes and cameras are bound by fundamental physical limits—rooted not in technology, but in the very nature of light itself. This means pinpointing an object’s exact location with infinite precision is physically impossible. Figure 1 shows edge of what physics allows.

Given this, the international research team set out to explore the ultimate precision optical systems can achieve. More importantly, they asked: can AI help us reach that theoretical limit? Their answer: yes—nearly. By training neural networks on complex light patterns, they’ve developed a method that comes astonishingly close to this boundary [1]. The technique is now being considered for real-world applications in fields like medical imaging.

Imaging Through Complexity

“Imagine trying to spot a small object behind a fogged-up or irregular glass pane,” explains Prof. Stefan Rotter of TU Wien. “Instead of a clear image, you see a chaotic pattern of light and dark patches. The question becomes: how accurately can we locate the object based on this scattered pattern—and what’s the best accuracy physics allows?”

This question isn’t just theoretical—it’s critical in real-world situations like biomedical imaging, where light is scattered unpredictably by tissue. Even when the data appears chaotic, can some meaningful information still be extracted?

To answer this, the team used a key concept from information theory: Fisher information, which quantifies how much data a signal carries about an unknown variable, such as position. If Fisher information is low, no algorithm, no matter how advanced, can provide accurate estimates. Using this concept, the researchers calculated the absolute upper bound of possible measurement precision for different scenarios.

Teaching AI to Decode Light Chaos

While TU Wien tackled the theory, researchers at Grenoble and Glasgow designed a real-world experiment. A laser beam was directed at a reflective object behind a turbid fluid, producing images of heavily distorted light patterns. The level of image distortion varied with the cloudiness of the fluid.

“To the human eye, these images look completely random,” says Maximilian Weimar, co-author from TU Wien. “But if you train a neural network with thousands of these images—each labeled with the object’s known position—the network begins to recognize subtle correlations.”After training, the AI could infer object positions with surprising accuracy, even from unseen patterns.

Approaching the Unbreakable Limit

Most impressively, the AI's accuracy came remarkably close to the theoretical limit set by Fisher information. “Our AI system isn’t just good—it’s nearly optimal,” says Rotter. “It achieves precision that is virtually at the edge of what physics allows.”

A Future for Ultra-Precise AI Imaging

This breakthrough has the potential to revolutionize optical measurement in fields like diagnostics, materials science, and quantum sensing. The researchers now plan to collaborate with applied physicists and medical experts to bring these AI-driven techniques into practical use.

As this study shows, combining physics with artificial intelligence doesn’t just make our tools smarter—it lets us stretch them to the very edge of possibility.

Reference:

1. https://scitechdaily.com/ai-pushes-imaging-to-the-absolute-brink-of-physical-limits/

Cite this article:

Keerthana S (2025), AI Takes Imaging to the Edge of What Physics Allows, AnaTechMaz, pp.759