Optimal-Transport Contrastive Learning for Centrality Determination in Heavy-Ion Collisions

summary

The gist

Determining centrality of heavy-ion collisions with high resolution is essential for event-by-event analyses…

In short

The method uses optimal transport distance to compare event momentum distributions against landmark events to determine collision centrality. It employs contrastive learning to create a compact event embedding, successfully clustering events by participant number (Npart). While effective at resolving centrality classes better than traditional slicing, applying realistic kinematic cuts causes a loss of distinct features.

Key concepts

Optimal Transport (Wasserstein) Distance
This is a mathematical measure used to quantify the minimum average distance required to move the particles of one event's momentum distribution onto another event's distribution. It helps determine how similar two events are in terms of their final particle states, acting as a sophisticated similarity metric for centrality.
Contrastive Learning
This is a machine learning technique used to train a network by showing it pairs of data where similar items (positives) are pulled closer together while dissimilar items (negatives) are pushed apart. Here, it trains the network to embed events such that events with similar physical characteristics, like Npart, end up close in the embedding space.
Centrality Resolution
This metric measures how well the method can distinguish between different collision classes (like central vs. peripheral) within each centrality bin. A higher resolution means the distribution of participant numbers (Npart) is more sharply defined within that class, indicating better separation of physical effects.
Event Embedding
This is a compact, low-dimensional representation created by the neural network that captures the essential features of an entire event's final particle momentum distribution. This embedding allows for efficient comparison and clustering of events based on their underlying collision geometry.

Terminology used across episodes

This episode discusses

The paper

Optimal-Transport Contrastive Learning for Centrality Determination in Heavy-Ion Collisions · Read on arXiv

Shu-Yun Yang, *, Yu Zhang, †

School of Artificial Intelligence in Medicine, Guilin Medical University · Guangxi Key Laboratory of Nuclear Physics and Technology, Guangxi Normal University

Determining centrality of heavy-ion collisions with high resolution is essential for event-by-event analyses, in particular for net-proton fluctuation measurements aimed at locating the QCD critical point. Supervised machine-learning approaches rely on labels generated by model and therefore inherit the model dependence of the training target. In this work, we propose an unsupervised method to determine collision centrality based on optimal transport. Each collision is represented by its 1-Wasserstein distances to a fixed set of landmark events chosen to cover the measured multiplicity range. A point-cloud network trained by contrastive learning then compresses these distances into a compact event embedding, which is partitioned into equal-size centrality classes. We find that this approach groups events with similar N part and yields a better centrality resolution, quantified by the normalized width of the N part distribution, than the conventional method of cutting on particle multiplicity. The method, however, encounters a bottleneck once realistic kinematic cuts are applied to the particles. We attribute this degraded performance to the loss of distinct features between central and peripheral collisions because of applying kinematic cuts. The application of optimal transport to centrality determination in experimental data, remains an open problem that needs further study.

Transcript

Introduction to the show: ident: High Energy Physics: Phenomenology. The day in particle phenomenology: models, signatures and constraints.

Kai: We're looking at the paper: "Optimal-Transport Contrastive Learning for Centrality Determination in Heavy-Ion Collisions".

Mira: Here's the gist of the paper. Determining centrality of heavy-ion collisions with high resolution is essential for event-by-event analyses, in particular for net-proton fluctuation measurements aimed at locating the QCD critical point The proposed unsupervised method uses optimal transport to determine collision centrality based on final particle momentum distributions The core of the approach involves comparing the full final-state momentum distributions of events using the optimal-transport (Wasserstein) distance to a fixed set of landmark events Each event is represented by its 1-Wasserstein distances to these landmarks, and a point-cloud network trained by contrastive learning compresses these distances into a compact event embedding partitioned into equal-size centrality classes The Wasserstein distance is defined as the minimum average distance by which the particles of one event are moved to match those of the other, calculated numerically using the Hungarian algorithm Training Signal Construction The contrastive learning objective is designed such that "the embedding reproduces the geometry of the transport feature space" Within each minibatch, "the k = 16 nearest neighbors of each event in F space are taken as positives" to minimize the InfoNCE loss The network is trained on a small amount of UrQMD events, specifically "the first 2000 events with Npart > 0 are used as the training sample" Performance and Comparison The study compares the OT-contrastive learning method against conventional methods like Nch slicing, and finds that "the 'OT, all' which utilize all particles has best performance" The OT-contrastive method successfully clusters events with similar Npart into the same centrality class However, when realistic kinematic cuts are applied, "the distinct features in momentum distribution between central and peripheral collisions are lost" Centrality Resolution Analysis The performance is quantified by the centrality resolution, defined as "the ratio of the standard deviation to the mean (σ(Npart)⟨Npart⟩) of the Npart distribution in each centrality class" The "OT, all" method yields a better centrality resolution than conventional methods Conversely, the "OT, charged, η" result has the worst centrality resolution and even worse than the conventional method of cutting on Nch distribution “Nch slicing” The authors attribute this degradation to "the loss of distinct features between central and peripheral collisions because of applying kinematic cuts" The method demonstrates that the OT-contrastive learning approach can still extract some information about collision geometry even when using only charged particles with a realistic η cut Further studies are needed to apply optimal transport to centrality determination in heavy-ion collisions The bottleneck encountered when applying kinematic cuts suggests that future work should focus on mitigating the loss of distinct features between collision classes ACKNOWLEDGEMENT Shu-Yun Yang is supported by the Guangxi Young and Mid-Career University Teachers’ (Research) Basic Capacity Enhancement Project (No.

Kai: 2025KY0533).

Mira: Yu Zhang was supported by the Guangxi Natural Science Foundation (No.

Kai: 2025GXNSFBA069067).

Paper summary: Kai: So we're looking at this paper titled Optimal-Transport Contrastive Learning for Centrality Determination in Heavy-Ion Collisions by Shu-Yun Yang and Yu Zhang. The main idea is using an unsupervised method to figure out collision centrality based on particle momentum distributions.

Mira: It claims that each collision is represented by its one Wasserstein distance to a set of fixed landmark events, and then a point cloud network trained with contrastive learning compresses those distances into a compact event embedding, which gets split into equal-size centrality classes.

Lev: So it's using optimal transport to map the shape of the momentum cloud rather than just counting particles, which is interesting because that suggests it’s trying to capture more detail about the collision geometry than just looking at multiplicity or transverse momentum.

Kai: Exactly. The authors are proposing this method because determining centrality with high resolution is really necessary for event-by-event analyses, especially when we're trying to measure net-proton fluctuations near the QCD critical point.

Mira: They suggest that supervised machine learning approaches have a problem because they need labels generated by a model, and that introduces model dependence into the training target. So this unsupervised method aims to avoid that issue by using optimal transport as its core feature characterization mechanism.

Lev: From an engineering standpoint, the concept of using the one Wasserstein distance to map out the features is cool, but we have to consider how much complexity that network actually adds when you put it on real hardware.

Kai: The paper does present some results showing that this approach groups events with similar net participant nucleon numbers, or Npart. It also mentions that they found improvements by a factor of two compared to classical techniques based on particle multiplicity and transverse momentum, pT.

Mira: But they also included a warning about model bias from training on a transport model, so the authors propose doing a cross-model test to estimate how much of that bias is actually there.

Lev: That makes sense. If you train your correction method on UrQMD events, you have to be careful that it doesn't just learn the artifacts of that specific model rather than the physics we want to measure.

Paper summary: Kai: The paper also mentions previous work, noting that machine learning techniques for centrality determination moved from raw detector-level input, like CBM hits and reconstructed tracks from full simulations, to higher accuracy on simulated data using a combined CNN and LightGBM.

Mira: And they point out that there are other attempts using transfer learning techniques to fine-tune pre-trained models to predict Npart. The work in this paper is situated within that broader area of applying machine learning for impact parameter determination, like the work done by Bass et al. in one thousand nine hundred ninety-four and later neural network approaches <ref:2611.07465#pg1>.

Lev: It’s interesting that they tested different scenarios, like using all outgoing particles versus just charged particles or charged particles with a specific eta cut corresponding to STAR fixed-target and collider data acceptance.

Kai: The performance is quantified by the centrality resolution, which they define as the ratio of the standard deviation to the mean of the Npart distribution in each centrality class. And they found that their method, specifically using all particles, yields a better centrality resolution than conventional methods like Nch slicing.

Mira: Conversely, they noted that when realistic kinematic cuts are applied—like those for charged particles with eta cut—the distinct features in momentum distribution between central and peripheral collisions get lost.

Lev: So the limitation they flag is that applying those kinematic cuts seems to lose the physical separation between the collision classes when using charged particles. That’s a key caveat we need to keep in mind when thinking about real experiments.

Kai: The paper also highlights that while the OT-contrastive learning approach successfully clusters events with similar Npart, it still struggles when those realistic kinematic cuts are introduced because of this loss of distinct features.

Mira: The authors attribute this degradation directly to the loss of distinct features between central and peripheral collisions because of applying kinematic cuts. This is a crucial piece of information for anyone trying to use these results for net-proton fluctuation measurements aimed at locating the QCD critical point.

Lev: So what this means for running this on real hardware is that if we want good resolution, we probably need to stick closer to the full particle information, like they did in the 'OT, all' scenario.

Paper summary: Kai: The authors also compare their method against other approaches, finding that while their 'OT, all' method has the best performance overall according to them, other methods like 'OT charged eta' have the worst centrality resolution and even worse than conventional Nch slicing.

Mira: That suggests that while the optimal transport framework is powerful for feature extraction, the choice of which features to use—all particles versus just charged particles—really dictates how good your final centrality classes are.

Lev: It sounds like future work needs to focus on mitigating this loss of distinct features when applying kinematic cuts, because right now, that seems like a major bottleneck for using this in practice.

Kai: So to wrap up the main points of Optimal-Transport Contrastive Learning for Centrality Determination in Heavy-Ion Collisions: the thesis is that optimal transport can characterize the shape of momentum clouds rather than just particle counts, and contrastive learning compresses those distances into embeddings partitioned into centrality classes.

Mira: It claims improvement over classical techniques by a factor of two in some cases, but it also carries the caveat that training on a transport model introduces a model bias, which they suggest testing with cross-model tests.

Lev: And the implication for us is that this unsupervised method has shown promise for extracting event geometry from raw particle data, but its ability to resolve centrality depends heavily on whether you use all particles or just charged particles with specific kinematic cuts.

Kai: The authors are using this framework because determining collision centrality with high resolution is essential for event-by-event analyses, particularly for net-proton fluctuation measurements aimed at locating the QCD critical point.

Mira: In conclusion, the paper shows that this optimal transport approach can still extract some information about collision geometry even when only using charged particles with a realistic eta cut.

Lev: The bottleneck they encountered when applying kinematic cuts suggests that future work should focus on mitigating the loss of distinct features between central and peripheral collisions.

Kai: So, for anyone listening who wants to know what this means, it’s that while the math is sophisticated using optimal transport and contrastive learning, you have to be careful about what you cut out—the momentum distributions—because those cuts seem to destroy the very differences they are trying to measure.

Conclusion: Kai: So we're wrapping up on this paper, "Optimal-Transport Contrastive Learning for Centrality Determination in Heavy-Ion Collisions."

Mira: It’s essentially showing how you can use optimal transport to figure out which collision event is which, based purely on how the particles are moving.

Lev: The core idea is using these Wasserstein distances to map the shape of the final particle momentum distribution and then using contrastive learning to compress that shape into a compact event embedding.

Kai: It's about moving away from just counting things and capturing the actual geometry of the collision in a way that machine learning can understand.

Mira: The authors are testing this against standard methods, like those based on particle multiplicity or transverse momentum slices. They found that when using all particles, this optimal transport approach gives a better resolution for how we classify events by their participant number, Npart.

Lev: But they also flagged a major issue—when you apply those realistic kinematic cuts to only charged particles with an eta cut, the distinct features between central and peripheral collisions get lost.

Kai: That loss of detail is significant because if we can't tell the difference between a central and a peripheral collision at the level needed for net-proton fluctuations, it hurts our ability to locate things like the QCD critical point.

Mira: So while this method works well when you look at all particles, that limitation with charged particles means we still have a problem if we want to use these results in actual experiments.

Lev: The paper also points out that while this method can extract some geometry even with those cuts, the bottleneck they found is exactly how to mitigate the loss of those distinct features when applying kinematic cuts.

Kai: It suggests that for future work, we really need to focus on developing methods that can preserve the physics information even when you're selecting only a subset of particles.

Mira: That opens up a new avenue for theory and simulation—if we can solve that feature loss problem, it could make event-by-event fluctuation measurements much more precise.

Lev: Right, so the challenge isn't just building the network; it’s making sure the input data you feed into it still carries all the necessary physics to make a good decision about centrality.

More episodes

← Home