Jet Origin Identification (JOI) in Physics

• Jet Origin Identification (JOI) is the process of inferring the source of a jet—from parton-level production in colliders to multi-phase outflows in astrophysics—using indirect observables.
• JOI employs diverse methodologies such as enhanced flavor tagging, graph-based networks, transformer models, and lifetime-based decay analyses to distinguish jet origins.
• JOI advances precision measurements from setting Higgs decay limits in collider experiments to resolving jet structures and launch regions in astrophysical observations.

Jet Origin Identification (JOI) denotes the inference problem of determining the physical origin of a jet from observables measured after propagation, fragmentation, or radiative processing. In collider physics, JOI is the procedure to determine which colored Standard Model particle originally produced a jet, extending conventional flavor tagging and quark–gluon discrimination to origin-level classification of quarks, anti-quarks, gluons, and heavy-particle jets (Liang et al., 2023). In astrophysical contexts represented in the same literature label, JOI refers to identifying which physical component, launch region, or dynamical mechanism a jet tracer corresponds to, for example distinguishing a fast collimated jet, a slower wide-angle wind, a broader outflow cavity, an internal working surface, or a large-scale inverse-Compton-emitting region (Dishoeck et al., 12 May 2025). Across these uses, the common problem is origin inference from indirect signatures rather than direct observation of the launch point or initiating parton.

1. Conceptual scope and problem statement

In the collider formulation, JOI generalizes several established tasks. The 11-category formulation proposed for an electron–positron Higgs factory classifies jets as $b,\bar b,c,\bar c,s,\bar s,u,\bar u,d,\bar d,g$, thereby combining jet flavor tagging, quark–gluon separation, and jet charge determination within a single origin-level framework (Liang et al., 2023). A related multiclass setting in hadronic collisions classifies light-flavor quark jets, gluon jets, $W$-boson jets, $Z$-boson jets, and top-quark jets, emphasizing the distinction between ordinary QCD cascades and boosted heavy-particle decays (Moreno et al., 2019). In heavy-ion and hadron-collider analyses, JOI is also used in narrower flavor-specific senses, such as identifying bottom-quark-initiated jets through displaced vertices and impact-parameter information [(Nguyen, 2012); (Collaboration et al., 2010)].

The astrophysical use of the term is broader and more diagnostic. In embedded protostars, JOI is described as identifying which physical component a given line traces, how close to the launch region it originates, and whether the emission indicates a fast collimated jet, a slower wide-angle wind, or a broader outflow cavity and shocked envelope (Dishoeck et al., 12 May 2025). In blazar and kiloparsec-jet studies, JOI becomes a problem of localizing the emitting region, for example deciding whether very-high-energy $\gamma$-rays arise in the compact core or in a $>100$ kpc extended jet (Zacharias et al., 2016), or whether optical, radio, and $\gamma$-ray variability is produced by co-spatial or downstream zones in a twisting relativistic flow (Raiteri et al., 2024). In laboratory and nozzle flows, the same logic appears as the attribution of a tonal jet response to a global resonant mode rather than to local shock breathing (Bakulu et al., 2020).

This breadth suggests that JOI is not a single standardized task across all jet sciences. A plausible implication is that the term names a shared inverse problem—origin attribution from downstream observables—while the relevant state spaces, observables, and validation standards remain domain-specific.

2. Collider JOI: partonic origin, heavy-particle tagging, and quark–gluon separation

A formal collider realization of JOI was given for $e^+e^- \to \nu\bar{\nu}H,\ H\to jj$ at $\sqrt{s}=240\ \text{GeV}$, using full Geant4 simulation of the CEPC baseline detector, $e^+e^-$-$k_t$ jet clustering, and a modified ParticleNet classifier (Liang et al., 2023). In the default particle-identification scenario, the resulting jet flavor tagging efficiencies are 92% for $W$0, 79% for $W$1, and 67% for $W$2, while gluon jets are identified with about 67% efficiency; charge flip rates are 19% for $W$3, 7% for $W$4, and 17% for $W$5, with $W$6 and $W$7 tagging efficiencies of 37%–41% and charge flip rates of 13%–24% (Liang et al., 2023). The same study emphasizes that $W$8-tagging improves from 47% with only lepton ID to 67% with charged-hadron ID, and to 74% when neutral kaons are added, while $W$9 information does not improve charge separation (Liang et al., 2023).

For boosted heavy-particle versus gluon discrimination, zest was introduced as a hadron-level observable defined by

$Z$0

with $Z$1 measured with respect to the jet axis (Budhraja et al., 2017). Zest was proposed for distinguishing gluon jets from jets originating from top quarks and vector bosons, with claimed properties including boost invariance along the jet axis, insensitivity to the inclusion or exclusion of a few soft particles, stability against global color flow of partons, and, for gluon jets, a narrow distribution nearly independent of jet mass (Budhraja et al., 2017). The reported qualitative pattern is that gluon jets peak around $Z$2, while $Z$3, $Z$4, and top jets populate higher zest values (Budhraja et al., 2017).

A subsequent generalization defined

$Z$5

and showed that $Z$6 yields $Z$7, approximately inverse multiplicity for $Z$8, while $Z$9 approaches the leading-particle limit $\gamma$0 (Budhraja et al., 2020). In that study, the optimal discrimination occurred for roughly $\gamma$1, and a zest cut was reported to remove about 90% of gluons while retaining roughly 80–90% of the heavy-particle signal in the cases studied (Budhraja et al., 2020).

The paper pairing zest with boost-invariant broadening, or bib, used two-dimensional discrimination in the $\gamma$2 plane, where gluon jets cluster at small zest and small bib, while heavy-particle jets occupy larger zest and larger bib regions (Budhraja et al., 2017). Exclusion-zone cuts were reported to be slightly better than zest-only cuts for weak bosons when high signal efficiency is desired, whereas for top quarks the exclusion zones did not significantly improve over zest alone (Budhraja et al., 2017).

These formulations make clear that collider JOI includes both highly structured multiclass origin classification and more targeted vetoes, especially gluon rejection in heavy-particle searches.

3. Lifetime-based JOI and displaced-vertex methods for heavy flavor

Heavy-flavor JOI exploits the long lifetime of $\gamma$3-flavored hadrons. In the D0 experiment, the basic physical handle is that with an average lifetime of about 1.5 ps, a $\gamma$4 hadron can travel of order millimeters before decay, producing charged tracks displaced from the primary interaction vertex (Collaboration et al., 2010). D0 implemented three complementary lifetime-based taggers—Secondary Vertex Tagger (SVT), Jet Lifetime Probability tagger (JLIP), and Counting Signed Impact Parameter tagger (CSIP)—and then combined them in a neural-network tagger (Collaboration et al., 2010). The signed transverse impact parameter $\gamma$5 and its significance,

$\gamma$6

provide the central low-level observables; light-flavor jets show an approximately symmetric signed-$\gamma$7 distribution around zero, whereas $\gamma$8 decays produce a strong positive tail (Collaboration et al., 2010). JLIP constructs a per-jet probability from track probabilities derived from the negative-$\gamma$9 side of the impact-parameter resolution function, while SVT reconstructs an explicit displaced secondary vertex and uses variables such as vertex mass and decay-length significance $>100$0 (Collaboration et al., 2010). The final D0 neural network used seven inputs and reported efficiency gains of about 20–50% at low fake rate, with a reduction of the fake rate by a factor of 2–3 compared with individual taggers (Collaboration et al., 2010).

CMS extended this style of JOI to the much denser PbPb environment, where the target was bottom-quark-initiated jets in heavy-ion collisions (Nguyen, 2012). The main discriminator was SSVHE (Simple Secondary Vertex High Efficiency), based on the flight-distance significance of the reconstructed secondary vertex, with JP (Jet Probability) used as a reference tagger based on impact-parameter significance (Nguyen, 2012). A central technical adaptation was to reconstruct jets first, use each jet as a seed, and restrict the tracker search region to a window around the jet axis for displaced-track reconstruction (Nguyen, 2012). In PbPb, the reported performance at about 50% b-jet efficiency gave light-jet rejection of roughly 100 and charm rejection of about 10, about a factor of 3 worse than in pp for light-jet rejection at the same efficiency (Nguyen, 2012).

CMS extracted the $>100$1-jet to inclusive jet ratio using the formula

$>100$2

where $>100$3 is the purity from secondary-vertex-mass fits and $>100$4 is the tagging efficiency (Nguyen, 2012). For 0–100% PbPb centrality, the reported $>100$5-jet fraction is approximately 2.9–3.5%, with no significant jet-$>100$6 dependence and absolute uncertainty of about 0.6–1.1% (Nguyen, 2012). The physics conclusion was cautious: within the current uncertainties, the data disfavor an extreme scenario in which $>100$7-jets experience no energy loss in PbPb collisions, but the uncertainties were too large to resolve detailed jet-$>100$8 dependence or sharply discriminate between flavor-dependent energy-loss models (Nguyen, 2012).

Within collider JOI, these lifetime-based methods remain distinct from substructure observables such as zest. They identify partonic origin through decay topology and displaced tracking rather than through the internal momentum-sharing pattern of prompt jet constituents.

4. Learned representations: interaction networks, transformers, and scaling behavior

JEDI-net cast JOI as learning on an unordered set of jet constituents connected by pairwise relations (Moreno et al., 2019). Jets were represented by up to 150 highest-$>100$9 particles, each with 16 features, and processed as a fully connected directed graph with $\gamma$0 edges (Moreno et al., 2019). The architecture defined an input matrix $\gamma$1, sender and receiver matrices $\gamma$2, an edge construction

$\gamma$3

message aggregation

$\gamma$4

and a post-interaction constituent representation

$\gamma$5

The study emphasized that this representation is permutation-insensitive, does not require image-like geometry or special sparse-input handling, and directly models pairwise interactions among constituents (Moreno et al., 2019).

On a five-class benchmark of light-flavor quark, gluon, $\gamma$6, $\gamma$7, and top jets, JEDI-net outperformed DNN, CNN, and GRU baselines at most operating points (Moreno et al., 2019). At false-positive rate $\gamma$8, true positive rates were 0.878 for gluons, 0.822 for light quarks, 0.938 for $\gamma$9, 0.910 for $e^+e^- \to \nu\bar{\nu}H,\ H\to jj$0, and 0.930 for top, all exceeding the corresponding CNN and GRU values and usually the DNN values (Moreno et al., 2019). At false-positive rate $e^+e^- \to \nu\bar{\nu}H,\ H\to jj$1, JEDI-net gave 0.485 for gluons and 0.769 for $e^+e^- \to \nu\bar{\nu}H,\ H\to jj$2, while for top the DNN slightly led with 0.651 versus 0.633 (Moreno et al., 2019). The model used relatively few parameters—33,625 for JEDI-net and 8,767 for the summed-$e^+e^- \to \nu\bar{\nu}H,\ H\to jj$3 variant—but was computationally expensive in FLOPs and inference time because of the fully connected graph (Moreno et al., 2019).

A distinct line of work asked whether a task-agnostic transformer could reach the performance of specialized JOI architectures on large numerical datasets (Wu et al., 2024). BBT-Neutron is a decoder-only transformer using binary or byte tokenization, patch embedding, multi-head causal self-attention, feed-forward layers, and Rotary Position Embeddings, with about 160 million parameters and no pretraining (Wu et al., 2024). For JOI, jets are serialized into bytes using per-particle attributes including $e^+e^- \to \nu\bar{\nu}H,\ H\to jj$4, $e^+e^- \to \nu\bar{\nu}H,\ H\to jj$5, $e^+e^- \to \nu\bar{\nu}H,\ H\to jj$6, $e^+e^- \to \nu\bar{\nu}H,\ H\to jj$7, impact-parameter observables, charge, and particle-type indicators such as isElectron, isMuon, isChargedKaon, isChargedPion, isProton, isNeutralHadron, and isPhoton (Wu et al., 2024). The tokenization workflow is summarized as

$e^+e^- \to \nu\bar{\nu}H,\ H\to jj$8

The CEPC JOI benchmark in that work used 11 labels—B-jet, B-bar-jet, C-jet, C-bar-jet, S-jet, S-bar-jet, U-jet, U-bar-jet, D-jet, D-bar-jet, G-jet—with training sets ranging from 100 events up to 10 million events, and an evaluation setting with 1 million jets per species split 60% train / 20% validation / 20% test (Wu et al., 2024). At the 10-million-statistics scale, BBT-Neutron, ParticleNet, and Particle Transformer were reported to show comparable performance in confusion matrices, flavor-tagging efficiency, and charge flip rate (Wu et al., 2024). The study highlighted an S-curve scaling pattern: below 10,000 events, BBT-Neutron was near random on jet charge; above that threshold its performance improved sharply; and at around 3,000,000 events its charge flip rate became comparable to ParticleNet and Particle Transformer (Wu et al., 2024).

These results mark a methodological divide within collider JOI. Specialized graph and particle-cloud architectures encode symmetry-aware inductive biases from the outset, whereas byte-tokenized transformers are more general and more data-hungry. This suggests that the representation question—sets and relations versus serialized numeric streams—is itself part of the JOI research problem.

5. Precision applications and limits in collider JOI

The principal precision application in the supplied literature is Higgs rare and exotic decay measurement at a future Higgs factory (Liang et al., 2023). Using JOI outputs for the two jets as inputs to a GBDT classifier, the study projected 95% confidence-level upper limits on branching ratios for rare decays $e^+e^- \to \nu\bar{\nu}H,\ H\to jj$9 and flavor-violating decays $\sqrt{s}=240\ \text{GeV}$0 at CEPC nominal luminosity of $\sqrt{s}=240\ \text{GeV}$1, corresponding to about 4 million Higgs bosons (Liang et al., 2023). The combined upper limits were reported as 0.75 × 10$\sqrt{s}=240\ \text{GeV}$2 for $\sqrt{s}=240\ \text{GeV}$3, 0.91 × 10$\sqrt{s}=240\ \text{GeV}$4 for $\sqrt{s}=240\ \text{GeV}$5, 0.95 × 10$\sqrt{s}=240\ \text{GeV}$6 for $\sqrt{s}=240\ \text{GeV}$7, 0.22 × 10$\sqrt{s}=240\ \text{GeV}$8 for $\sqrt{s}=240\ \text{GeV}$9, 0.23 × 10$e^+e^-$0 for $e^+e^-$1, 0.39 × 10$e^+e^-$2 for $e^+e^-$3, and 0.86 × 10$e^+e^-$4 for $e^+e^-$5 (Liang et al., 2023). The derived $e^+e^-$6 limit is approximately three times the Standard Model prediction quoted in the same paper, $e^+e^-$7 (Liang et al., 2023).

The same paper used the CL$e^+e^-$8 method at 95% confidence level, with both cut-and-count and shape-fit approaches, and found that the shape fit gave better sensitivity (Liang et al., 2023). For the $e^+e^-$9 channel in the $k_t$0 example, the optimal cut left 37 signal events and 5.1k background events, giving a signal-strength upper limit of 3.8, improved to 3.5 by fitting the score distribution and to 3.2 after combining with $k_t$1 and $k_t$2 (Liang et al., 2023).

Heavy-ion heavy-flavor JOI provides another precision application, though with a different target. The CMS PbPb study framed direct $k_t$3-jet identification as a route to flavor-resolved jet quenching, because comparing $k_t$4-jet suppression to inclusive jet suppression probes whether heavy quarks lose less energy than light partons and whether mass-dependent effects survive at high jet $k_t$5 (Nguyen, 2012). Within current uncertainties, the measurement gave an early constraint on the flavor dependence of parton energy loss but not a decisive model separation (Nguyen, 2012).

A common misconception is that JOI in collider physics is synonymous with $k_t$6-tagging. The literature here shows a broader structure: lifetime tagging is one important instance, but JOI also includes multiclass quark/anti-quark/gluon identification, boosted heavy-particle tagging, charge-sign inference, and gluon vetoing in substructure space (Liang et al., 2023, Budhraja et al., 2017).

6. Astrophysical JOI: launch regions, stratification, and emitting zones

In the protostellar literature, JOI is an exercise in tracer attribution and launch-region inference rather than parton classification. The JOYS program uses JWST/MIRI-MRS $k_t$7 integral-field spectroscopy with resolving power $k_t$8, mJy sensitivity, and sub-arcsecond imaging to spatially separate the central protostar, the jet axis, the cavity walls, and off-source knot or bow-shock positions (Dishoeck et al., 12 May 2025). Its most important JOI result is the repeatedly observed nested, stratified jet structure: an inner ionized core traced by [Fe II], surrounded by a molecular layer traced by higher-excitation H$k_t$9, with an even broader lower-excitation H$W$00 component tracing the wide-angle wind inside the cavity (Dishoeck et al., 12 May 2025). In this framework, refractory species such as Fe, Ni, and Co are interpreted as jet tracers linked to dust destruction, [S I] is an evolutionary diagnostic that follows the jet in Class 0 but becomes compact and on-source in Class I, and [Ne II] is a mixed diagnostic of jet shocks and photoionized emission (Dishoeck et al., 12 May 2025).

The JOYS+ survey sharpened this picture with a sample-based H$W$01 analysis (Francis et al., 15 Apr 2026). Low-$W$02 H$W$03 transitions trace extended wide-angle, low-velocity $W$04 winds within the contours of the low-velocity $W$05 sub-mm CO emission, while high-$W$06 transitions are associated with shocks and knots (Francis et al., 15 Apr 2026). In Class 0 sources with known high-velocity $W$07 molecular CO or SiO jets, higher H$W$08 velocities are found along the jet axis (Francis et al., 15 Apr 2026). The opening angle of the H$W$09 S(1) wind broadens from $W$10 to $W$11 from Class 0 to Class I, the warm component is $W$12 K and about two orders of magnitude more massive than the hot $W$13 K component, and the H$W$14 mass-loss rates decline by two orders of magnitude from the Class 0 to Class II stage (Francis et al., 15 Apr 2026). The authors interpret these trends as consistent with MHD disk wind models, while noting that the data do not uniquely exclude X-winds or strongly shocked or entrained interpretations for some components (Francis et al., 15 Apr 2026).

Individual systems show how JOI differentiates multiple ejection channels. In the Class I protobinary TMC1, TMC1-E powers a narrow H$W$15 outflow whose opening angle increases from S(8) to S(1), indicating a disk-wind origin, whereas TMC1-W powers a collimated [Fe II] and [Ni II] jet consistent with an energetic inner-disk flow (Tychoniec et al., 2024). The same study links an ALMA accretion streamer feeding TMC1-E from $W$16 au scales to the wide-angle molecular wind, while stronger H I recombination lines toward TMC1-W are associated with a collimated high-velocity jet within the innermost regions of the disk (Tychoniec et al., 2024). In HH 211, JWST shows an onion-like structure in which the atomic core is narrowest, H$W$17 0–0 S(7) forms a broader layer, H$W$18 0–0 S(1) is broader still, and warm H$W$19 mass flux, momentum, and momentum flux exceed the atomic values by up to a factor of ten, leading to the conclusion that the warm molecular jet is the primary dynamical driver of the outflow (Garatti et al., 2024).

JOI can also overturn simple tracer assignments. In 2MASS J16075796-2040087, KECK/HIRES spectro-astrometry showed that the lower-velocity forbidden-line emission previously classified as a disk-wind LVC has a jet-like centroid gradient, density and ionization fraction in the HVC range, and does not yield a distinct, clean MHD disk-wind component (Whelan et al., 2024). The preferred interpretation is that the HVC is a bona fide bipolar jet and the low-velocity emission is likely not a clean MHD disk wind but instead a slow jet component or a blend of slow jet and MHD wind emission (Whelan et al., 2024).

In Herbig–Haro jets, JOI can target the origin of internal structures rather than of the whole flow. For the HL Tau jet, high- and low-radial-velocity structures in each knot have essentially the same proper motion, supporting the interpretation that they are Mach disks and bow shocks within internal working surfaces produced by episodic velocity variations rather than instabilities in a stationary flow (Movsessian et al., 2012).

These examples show that astrophysical JOI is fundamentally multiphase. Atomic, molecular, ionic, and kinematic tracers are used to separate launch zones, shock structures, irradiation effects, and evolutionary state.

7. Resonant, relativistic, and extended-jet origin problems beyond protostars

Outside star formation, JOI in the supplied literature includes localization of emission and identification of global dynamical mechanisms. In AP Librae, the central claim is that the very-high-energy $W$20-ray emission originates in the $W$21 kpc extended jet rather than in the compact core (Zacharias et al., 2016). The evidence chain combines a hard X-ray jet spectrum with photon index $W$22, the failure of a standard one-zone core model to reproduce the TeV emission, and a successful fit using inverse Compton scattering of CMB photons in a weakly magnetized extended jet with $W$23, $W$24, $W$25, and electron spectral index $W$26 (Zacharias et al., 2016). Here JOI is an emitting-zone localization problem: the origin is the extended jet, not the core.

For BL Lacertae, the origin problem is geometric. The twisting jet model reconstructs time-dependent Doppler factors and viewing angles from optical, radio, and $W$27-ray light curves using

$W$28

with $W$29, $W$30, $W$31 in the optical, and $W$32 in the radio (Raiteri et al., 2024). The inferred optical and $W$33-ray regions are co-spatial, while the radio-emitting region follows the optical one by about 120 days, leading to a picture of a curved, twisting, inhomogeneous jet composed of a pair of emitting plasma filaments in a double-helix-like rotating structure (Raiteri et al., 2024). In this usage, JOI becomes reconstruction of relative spatial placement along a multi-filament relativistic jet.

In compressible nozzle flow, origin identification can refer to the dynamical source of a jet resonance. For an overexpanded jet in a truncated ideally contoured nozzle operating in free shock separation, synchronized wall-pressure and time-resolved stereo-PIV measurements found a pronounced tonal peak at $W$34, significant only in the first azimuthal mode $W$35, together with coherence between internal wall pressure and the external jet only at that tonal frequency (Bakulu et al., 2020). DDES and SPOD analysis showed that the resonant mode contains both downstream- and upstream-propagating waves, and that the side-load signature is directed by resonance of these coherent structures rather than by local shock breathing alone (Bakulu et al., 2020). The dominant wall-pressure SPOD mode near the tone carries nearly two orders of magnitude more energy than the second mode, and the decomposition attributes about 80% of the side-load amplitude to the shock-related component, about 13% to the upstream-propagating wave, and about 7% to the downstream-propagating Kelvin–Helmholtz-like wave (Bakulu et al., 2020).

Taken together, these cases show that JOI can denote origin attribution at very different levels: launch region, emitting zone, dynamical instability, or resonance closure mechanism. This suggests that the term is unified less by a common data modality than by a common epistemic structure: the origin is inferred from correlated morphology, kinematics, coherence, and model comparison.