Statistical validation of calorimeter inpainting with generative diffusion priors

In the high-stakes world of particle physics, scientists smash atoms together at speeds close to light to recreate the conditions of the early universe. To understand what happens in these collisions, they rely on massive detectors that act like giant, three-dimensional cameras, capturing the spray of energy and particles that fly out from the impact. One crucial component of these detectors is the calorimeter, a device designed to measure the energy of these particles with extreme precision. However, just like any complex machine, these detectors are not perfect. Over time, small sections can stop working, or "channels" can go dead, leaving gaps in the data. When a particle hits a dead spot, its energy signal is lost, creating a hole in the picture. If scientists cannot account for these missing pieces, their measurements of the collision become inaccurate, potentially skewing our understanding of fundamental physics.

The challenge is not simply to guess what was lost, but to reconstruct the missing information in a way that is statistically sound. Because many different patterns of energy could theoretically fit into a hole, a simple guess is not enough; scientists need a method that can generate a range of plausible possibilities and tell them how confident they should be in each one. This is where a new approach using advanced computer models comes in, offering a way to fill in the blanks of the universe's most energetic events with both accuracy and a clear measure of uncertainty.

Researchers at Stony Brook University have developed a sophisticated method to solve this problem of missing data in calorimeters. They focused on the Relativistic Heavy Ion Collider, where gold nuclei are smashed together, creating a chaotic environment filled with thousands of particles. In these collisions, the energy deposits form complex, swirling patterns. When a section of the detector fails, it leaves a square-shaped void in this pattern. The team used a type of artificial intelligence known as a generative diffusion model. Think of this model as a highly trained artist who has studied millions of perfect collision images. Instead of just guessing a single value to fill a hole, the model learns the underlying rules of how energy flows and spreads in these collisions. It can then generate a whole family of possible images that fit the surrounding data, effectively "painting" over the dead spots with physically realistic energy distributions.

The core of this study was to test whether these computer-generated fillings were not only visually convincing but also statistically reliable. The researchers compared several different algorithms that use this diffusion technology to see which one performed best. They treated the problem as a puzzle where the edges of the missing piece were known, and the goal was to reconstruct the center. They tested these methods against a baseline approach that simply filled the hole with an average value, a technique that often fails to capture the true complexity of the event. The team ran thousands of simulations, creating perfect "truth" images and then deliberately hiding parts of them to see how well the algorithms could recover the original data.

The results showed that the diffusion-based methods were far superior to the simple average-filling technique. The best-performing algorithms, particularly one called DDNM, were able to reconstruct the total energy of the collision with remarkable precision, keeping the overall energy balance almost exactly as it should be. More importantly, these methods did not introduce systematic errors or biases that would trick the scientists into thinking a collision was different than it really was. The reconstructed images maintained the correct spatial patterns, meaning the energy was placed in the right spots relative to the surrounding particles.

Perhaps the most significant finding was how well these models handled uncertainty. In science, knowing how sure you are about a measurement is just as important as the measurement itself. The researchers found that the best diffusion models provided a range of possible solutions that accurately reflected the true variability of the data. When they checked the statistical confidence of these predictions, the models proved to be well-calibrated, meaning their estimates of uncertainty matched the reality of the missing data. This is a crucial distinction because other methods might produce a single, clean-looking image but fail to tell the scientist how much they should trust that image. The diffusion models, by contrast, offered a complete picture of the possibilities, allowing physicists to carry forward a realistic sense of error into their final calculations.

The study also tested the robustness of these methods under different conditions. They examined whether the reconstruction quality changed depending on how crowded the collision was, simulating both sparse and extremely busy events. The models performed consistently well in both scenarios, showing that the technique is reliable regardless of the complexity of the event. Furthermore, they tested the size of the missing area, ranging from small gaps to large blocks covering a significant portion of the detector. Even as the missing area grew larger and the task became harder, the models adapted by naturally increasing their uncertainty estimates while still maintaining the correct average energy. This adaptability suggests that the method could handle real-world detector failures that vary over time and location.

Ultimately, this work establishes a new standard for validating how we recover lost information in high-energy physics. It moves beyond simply asking if a reconstructed image looks right, to asking if the statistical properties of that reconstruction are trustworthy. By proving that these generative models can faithfully reproduce missing detector signals and provide reliable uncertainty estimates, the researchers have opened the door for more precise measurements in future experiments. This approach ensures that when scientists look at the data from the most energetic collisions in the universe, they are not just seeing a filled-in picture, but a scientifically rigorous reconstruction that honors the true nature of the missing information.