⚛️ phenomenology

Too good to go: Upcycling Phase-Space Points for Multijet Processes

This paper introduces a training strategy that leverages the nested structure of phase spaces to efficiently adapt high-dimensional samplers for multijet processes by augmenting existing $N$ -particle datasets to initialize $(N+1)$ -particle sampling, thereby significantly reducing computational costs while improving performance over current benchmarks.

Original authors: Konrad Helms, Timo Janßen, Steffen Schumann

Published 2026-08-28

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Konrad Helms, Timo Janßen, Steffen Schumann

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

High-energy physics is the science of smashing particles together at speeds close to that of light to see what they are made of and how they interact. To understand these collisions, scientists rely on powerful computer programs called event generators. These programs act as virtual laboratories, simulating billions of potential collisions to predict what a real detector should see. The challenge lies in the sheer complexity of the data; when particles collide, they often produce a spray of smaller particles, and the number of possible ways these sprays can form is astronomical. To make sense of this, the computer must sample a vast, multi-dimensional space of possibilities, looking for the most likely outcomes. If the computer samples these possibilities inefficiently, it wastes immense computing power on unlikely scenarios, or worse, misses the rare but important ones entirely. This inefficiency becomes a major bottleneck when trying to study processes that produce many particles at once, such as jets of particles created alongside heavy particles like the top quark or the Z boson.

A team of researchers at the University of Göttingen has developed a new way to teach these computer programs how to sample these complex scenarios more efficiently. Instead of starting from scratch every time they need to simulate a collision with more particles, they found a way to reuse what the computer has already learned. In their study, they focused on processes where a core event, like the creation of a pair of top quarks or a Z boson, is accompanied by an increasing number of additional jets. Traditionally, to simulate a collision with five jets, the computer would have to learn the entire pattern of five jets from the ground up, a task that requires evaluating extremely expensive mathematical formulas millions of times. The researchers realized that the physics of these events has a nested structure: a collision with five jets is essentially a collision with four jets plus one extra jet. They proposed a method where the computer first masters the four-jet scenario and then uses that knowledge as a starting point to learn the five-jet scenario.

The researchers tested this idea by training a machine learning model to generate these particle collisions. They began by training the model on a dataset of four-jet events. Once the model was proficient at finding the most likely four-jet configurations, they took that dataset and simply added the coordinates for a fifth jet. They did this in two ways: either by filling the new jet's details with random guesses, or by copying the properties of an existing jet and shuffling them to create a new, plausible candidate. This created a massive, pre-sampled dataset for the five-jet scenario that already contained the correct patterns for the first four jets. They then used this augmented dataset to initialize the training for the five-jet model. Because the model started with a head start, it did not need to waste time learning the basic structure of the four-jet core; it only had to learn how the fifth jet interacts with the rest.

The results showed that this approach dramatically reduced the computational cost. In simulations of top-quark pair production with five accompanying jets, the new method reached the same level of accuracy as the traditional method while using only about 20 percent of the expensive calculations required to evaluate the underlying physics formulas. For the Z boson production with five jets, the savings were even more pronounced, cutting the cost by nearly 80 percent. Furthermore, the models trained with this shortcut did not just save time; they actually performed better. They produced more accurate predictions and generated events that were easier to analyze than those from the standard training methods. The researchers confirmed that this efficiency gain was not just because they started with a smaller dataset, but because the quality of the initial data was superior. By reusing the learned structure of the lower-multiplicity events, the model converged on the correct answer much faster.

This technique is not limited to machine learning. The authors noted that the same logic could be applied to older, traditional algorithms used in physics, suggesting a broad utility for the method. The core insight is that in complex physical systems, the rules governing a smaller system often remain valid when the system grows larger. By respecting this hierarchy and carrying forward the knowledge gained from simpler cases, scientists can bypass the most expensive parts of the learning process. The study demonstrates that for high-multiplicity processes, which are crucial for testing the limits of our understanding of the universe, there is a smarter way to build the virtual laboratories that help us explore them. The method ensures that the final results remain mathematically exact and unbiased, while the path to getting there becomes significantly shorter and less costly.

Technical Summary: MCNET-26-24

Problem Statement
Monte Carlo event generators face a significant computational bottleneck when simulating high-multiplicity final states in high-energy physics, such as multijet processes at the Large Hadron Collider (LHC). The evaluation of scattering matrix elements for these complex processes is computationally expensive. Consequently, training efficient samplers (e.g., via importance sampling or machine learning) requires a large number of matrix-element evaluations to adapt the sampling distribution to the target integrand. Traditional approaches train samplers for each jet multiplicity ( $N$ ) independently, often starting from uniform sampling. This independent training is particularly inefficient for high multiplicities where the phase space is high-dimensional and the matrix-element evaluation cost scales steeply.

Methodology
The authors propose a training strategy termed "Phase-Space Augmentation" that exploits the nested structure of phase spaces. The core idea is to transfer knowledge from a well-trained sampler for an $N$ -particle process to initialize the sampler for an $(N+1)$ -particle process.

Factorised Phase-Space Mapping: The method relies on the CHILI phase-space parameterisation, which factorises the $(N+1)$ -particle phase space into an $N$ -particle component and a single-particle component: $d\Phi_{N+1} \sim d\Phi_N \times d\Phi_1$ . This parameterisation ensures that the coordinates for the first $N$ particles retain their meaning and kinematic ranges when an additional particle is added.
Data Augmentation: Instead of generating new training data for the $(N+1)$ $(N + 1)$ -particle process from scratch, the authors take a dataset of phase-space points generated by the trained $N$ $N$ -particle sampler. They augment these points by appending three new kinematic coordinates (transverse momentum squared, rapidity, and azimuthal angle) corresponding to the additional particle.
- Uninformed Augmentation: The new coordinates are filled with uniform noise.
- Informed Augmentation: The new coordinates are filled by copying and shuffling coordinates from existing particles in the $N$ -particle dataset to preserve marginal distributions.
Helicity Conditioning: Since the number of non-vanishing helicity states doubles when adding a gluon, the augmented dataset is expanded by duplicating each event with two distinct helicity labels (one for the original configuration, one for the flipped new gluon helicity). The weights for these new labels are initialized by splitting the original weights.
Training Strategy: The augmented dataset serves as the initial proposal distribution for the iterative training of a Continuous Normalising Flow (CNF) using Conditional Flow Matching (CFM). Because the augmented points are only used as an initial proposal and not as the final target distribution, the method remains unbiased; exact event weighting is applied during the Monte Carlo integration to correct for any discrepancies. The authors note that this approach is agnostic to the sampler type and could theoretically apply to algorithms like VEGAS.

Key Contributions

Cross-Multiplicity Knowledge Transfer: The paper introduces a mechanism to reuse the learned transport maps and approximate distributions from lower-multiplicity processes to bootstrap training for higher-multiplicity processes.
Reduction in Computational Cost: By initializing the training of high-multiplicity samplers with augmented data from lower multiplicities, the method drastically reduces the number of expensive matrix-element evaluations required during the initial adaptation phase.
Improved Performance: The strategy not only reduces cost but also yields samplers with superior final performance (higher unweighting efficiency and lower integration error) compared to standard training, particularly for high-dimensional phase spaces.

Results
The authors tested the method on two processes at $\sqrt{s} = 13$ TeV: Drell–Yan production ( $d\bar{d} \to e^+e^- + ng$ ) and top-pair production ( $gg \to t\bar{t} + ng$ ) with up to $n=5$ gluons.

Efficiency Gains: For the $n=5$ gluon channel (19-dimensional phase space), the informed augmentation strategy reached the peak unweighting efficiency ( $\epsilon_{1\%}$ ) of a standard benchmark trained on 100M uniform events while using only ~22% of the actual training data (matrix-element evaluations).
Cost Reduction: The method reduced the computational cost of matrix-element evaluations by nearly 80% for the Drell–Yan 5-gluon process and showed similar significant reductions for top-pair production.
Superiority over Baselines: The augmented models consistently outperformed both a standard benchmark trained on 100M events (ST100) and a benchmark trained on a reduced dataset of 20M events (ST20). The advantage was most pronounced for higher multiplicities ( $n \ge 4$ ), where the fraction of phase-space dimensions inherited from the lower-multiplicity model is largest.
Convergence: Models trained with augmentation converged faster and reached higher performance plateaus with fewer iterations compared to standard training.

Significance and Claims
The paper claims that this augmentation strategy offers a "simple yet effective" way to reduce the cost of training phase-space generators for high-multiplicity final states. The significance lies in the ability to scale Monte Carlo event generation to higher jet multiplicities without a prohibitive increase in computational resources. The authors emphasize that:

The method preserves the unbiased nature of Monte Carlo estimates through exact event weighting.
It is applicable to various sampling methods, including machine learning-based flows and traditional algorithms like VEGAS.
The approach is particularly beneficial for "stacks" of processes with increasing jet multiplicity, where the speed-ups accumulate.
The framework is versatile and could potentially be extended to other applications, such as Next-to-Leading Order (NLO) calculations where real-emission degrees of freedom are added to a Born-level phase space.

The authors conclude that the quality of the initial data (derived from lower-dimensional, well-sampled projections) outweighs the sheer quantity of data in the early stages of training, allowing for more efficient exploration of complex, high-dimensional integrands.

More like this