Parnassus for the CLD Detector: A Generative Machine-Learning Surrogate for Detector Simulation and Reconstruction at the FCC-ee

The future of particle physics depends on a delicate balance between imagination and calculation. Scientists at the next generation of particle colliders, such as the proposed Future Circular Collider, plan to smash electrons and positrons together with unprecedented precision to study the fundamental building blocks of the universe. To understand what happens in these collisions, researchers rely on massive computer simulations that act as a virtual laboratory. These simulations must track how particles interact with a detector, a complex machine filled with sensors, magnets, and layers of material. The problem is that running these simulations is incredibly slow and computationally expensive. A single collision event can take minutes or even hours to process through the full chain of physics and engineering models. When scientists need to study billions of events to find rare phenomena, this speed becomes a bottleneck, making it difficult to design the detectors or plan the experiments effectively.

To solve this, researchers have turned to a new kind of artificial intelligence that learns to mimic the behavior of these complex machines. Instead of calculating every single interaction from scratch, these systems are trained on existing simulation data to predict the outcome directly. This approach, known as a surrogate model, acts like a highly skilled shortcut. It does not replace the laws of physics but rather learns the pattern of how a real detector would respond to a specific set of particles. By doing so, it can generate results in a fraction of the time, allowing scientists to test ideas and design experiments much faster than before. The challenge, however, is ensuring that this shortcut is accurate enough to be trusted. If the shortcut misses subtle details, it could lead to wrong conclusions about the nature of matter.

A team of physicists has now demonstrated a powerful new version of this shortcut tailored for a specific detector design called CLD, which is being developed for the Future Circular Collider. They built a system named Parnassus, which uses a type of generative machine learning to recreate the output of a full, high-fidelity simulation. In their work, they focused on a common type of collision where an electron and a positron annihilate to create a Z boson, which then decays into pairs of quarks. These quarks transform into sprays of particles called jets. The researchers trained their model on millions of simulated events where the particles were passed through a detailed digital version of the CLD detector and a sophisticated reconstruction algorithm that groups sensor signals into identifiable particles. The goal was to see if the AI could reproduce the final list of particles, their energies, and their paths with the same precision as the slow, traditional method.

The results show that the Parnassus model succeeds remarkably well. When the researchers compared the AI-generated events to the gold-standard simulations, the two matched almost perfectly across a wide range of measurements. The model correctly reproduced the energy and momentum of individual particles, the types of particles produced, and the complex patterns of how they are distributed in space. Crucially, it also captured the subtle details of how particles travel through the detector, including the tiny deviations in their paths that occur when they originate from short-lived heavy particles. This level of detail is essential for identifying specific types of matter, such as bottom quarks, which are key to many physics studies. The model also performed well when tested on particle types it had not seen during training, suggesting it has learned the underlying physics rather than just memorizing the data.

In contrast, the team compared their new model against a widely used, faster simulation tool called Delphes, which relies on simplified, hand-tuned formulas. While Delphes is fast, the study found that it failed to capture many of the complex correlations and distributions that the full simulation and the new AI model got right. For instance, Delphes struggled to reproduce the correct shapes of particle energy distributions and the detailed structure of jets. The new AI model, however, maintained a high level of fidelity, matching the full simulation so closely that the differences were often smaller than the statistical noise expected in real data. This indicates that the AI has learned the intricate relationship between the initial collision and the final detector response, including the messy, non-linear effects that simple formulas often miss.

Beyond just matching the data, the researchers tested whether the AI-generated events could be used for real physics analysis. They trained a separate machine learning tool to identify the flavor of the original quarks based on the particles produced by the AI. This tool performed just as well on the AI-generated data as it did on the full simulation data. This is a critical finding because it proves that the shortcut does not just look like the real thing; it behaves like the real thing for the complex tasks physicists actually need to perform. The model successfully preserved the ability to distinguish between different types of quarks, a task that requires precise knowledge of particle paths and energies.

The speed of this new method is staggering. While the full simulation and reconstruction chain takes a significant amount of time to process a single event, the Parnassus model can generate a complete event in about 1.2 milliseconds on a standard graphics card. This is more than a thousand times faster than the traditional method. Even on a standard computer processor, it runs hundreds of times faster. This massive gain in speed means that scientists can now generate the vast amounts of data needed for feasibility studies and detector design in a reasonable timeframe. It opens the door to exploring many more detector configurations and physics scenarios than was previously possible.

The researchers also showed that their framework is flexible. Because the model learns the relationship between the initial particles and the final detector output, it can be adapted to different reconstruction algorithms without needing to be rebuilt from scratch. They demonstrated this by showing how the same approach could be extended to a newer, machine-learning-based reconstruction method called HitPF. This adaptability suggests that the technique could become a standard tool for future collider projects, helping to bridge the gap between theoretical predictions and experimental reality.

The study confirms that generative machine learning can serve as a reliable and efficient surrogate for detector simulation. By training on high-quality data, the model learns to replicate the complex response of a detector with high precision, outperforming existing fast-simulation tools while being orders of magnitude faster. This achievement removes a major computational barrier for the design and study of future particle colliders. It allows physicists to focus on the science rather than the waiting time, ensuring that when the next generation of colliders comes online, the tools to understand them are ready. The work represents a significant step forward in using artificial intelligence to solve the practical challenges of high-energy physics, proving that speed and accuracy can coexist in the quest to understand the universe.