FENIX: A Multi-Disciplinary Research Label

• FENIX is a multi-domain label that refers to unrelated systems in forward spectrometry at RHIC, airborne hyperspectral sensing, in-network ML inference, and cosmological simulation.
• In high-energy physics and remote sensing, FENIX enables innovative pipelines—from detecting forward-angle jets and saturation phenomena to applying hyperspectral unmixing and CNN classification for simulated ground truth.
• In networking and cosmology, FENIX underpins hardware-software co-design for low-latency inference and detailed simulations of baryon-induced contraction and shape changes in dark-matter haloes.

Searching arXiv for the specified FENIX-related papers to ground the article and citations. FENIX is not a single research object but a label used for several unrelated systems across high-energy nuclear physics, remote sensing, programmable-network architectures, and cosmological simulation. In the cited arXiv literature, it denotes: the forward sPHENIX spectrometer, often colloquially called “FENIX” in the PHENIX/sPHENIX context; the SPECIM AisaFENIX 1K airborne hyperspectral sensor used to simulate ground truth for VENuS image classification; an FPGA-enhanced programmable-switch system for in-network DNN inference; and the Fenix cosmological hydrodynamical simulation suite used to study baryon-induced modification of dark-matter haloes (Seto, 2012, Faran et al., 2019, Gao et al., 20 Jul 2025, Cataldi et al., 2020).

1. Nomenclature and domain-specific usage

The same label appears in four distinct technical settings, with no common architecture, dataset, or scientific objective linking them. In the PHENIX/sPHENIX literature, “FENIX” refers to the forward sPHENIX spectrometer, or fsPHENIX. In remote sensing, FENIX is the airborne hyperspectral imaging system SPECIM AisaFENIX 1K. In networking, FENIX is a hybrid ASIC–FPGA platform for in-network ML. In galaxy-formation studies, Fenix is a cosmological hydrodynamical simulation suite.

Domain	Meaning of FENIX	Representative paper
RHIC detector instrumentation	Forward sPHENIX spectrometer (fsPHENIX)	(Seto, 2012)
Remote sensing	SPECIM AisaFENIX 1K airborne hyperspectral sensor	(Faran et al., 2019)
Programmable networks	FPGA-enhanced programmable-switch inference system	(Gao et al., 20 Jul 2025)
Cosmological simulation	Fenix hydrodynamical simulation suite	(Cataldi et al., 2020)

This multiplicity creates an immediate risk of category error. In particular, the term does not denote a single acronym, a unified software stack, or a transferable methodology across the four cited literatures. The only stable encyclopedic treatment is therefore disambiguating: each usage must be read within its disciplinary context.

2. FENIX in RHIC forward instrumentation

In the PHENIX/sPHENIX context, FENIX denotes the forward sPHENIX spectrometer, conceived as a forward-angle upgrade to extend sPHENIX coverage into the high-rapidity region. The design targets $1.2 < \eta < 4.0$, with two sections, $1.2\!-\!3.0$ and $3.0\!-\!4.0$, to measure hadrons, photons, electrons/muons for dileptons, and jets in $pp$, $p+A$, and $A+A$ collisions at RHIC. Its physics motivations are the study of high gluon densities and small-$x$ QCD in nuclei, together with transverse-spin and longitudinal-spin measurements enabled by polarized proton beams (Seto, 2012).

The kinematic rationale is standard forward-QCD logic. Pseudorapidity is defined by $\eta = -\ln \tan(\theta/2)$, so forward $\eta$ corresponds to small polar angle $\theta$ and, in $1.2\!-\!3.0$0, to small $1.2\!-\!3.0$1 in the nuclear target. For $1.2\!-\!3.0$2 hard scattering, the typical hard scale is $1.2\!-\!3.0$3, with

$1.2\!-\!3.0$4

The proposal explicitly frames the forward region as a laboratory for saturation-sensitive observables, using the parametric saturation-scale form

$1.2\!-\!3.0$5

The detector concept is organized into two forward sections. In Section I ($1.2\!-\!3.0$6), an extension or modification of the central solenoid provides a strong axial field. Charged-particle tracking is assigned to GEM stations, while silicon detectors near the interaction point provide precision vertexing for charm/beauty tagging and suppression of correlated heavy-flavor backgrounds in Drell–Yan. Hadron PID is assigned to a RICH detector. Calorimetry consists of a forward EMCal, reconfigured from existing PHENIX EMCal components, followed by HCAL; a preshower or charged-particle veto ahead of EMCal improves $1.2\!-\!3.0$7 identification and $1.2\!-\!3.0$8 separation. Muon identification behind HCAL is under feasibility study. In Section II ($1.2\!-\!3.0$9), the concept calls for a radial magnetic field with GEM tracking in the field region, together with a reconfiguration of the lead–tungstate Muon Piston Calorimeter for high-granularity EM measurement; hadronic calorimetry in this region remains under discussion.

The physics program is correspondingly broad. For small-$3.0\!-\!4.0$0 gluon dynamics, the proposal emphasizes forward hadron suppression, modification or loss of back-to-back azimuthal correlations, direct-$3.0\!-\!4.0$1+jet and dijet measurements, and Drell–Yan pair–hadron correlations. A central strategy is to use $3.0\!-\!4.0$2 to constrain the dipole gluon distribution $3.0\!-\!4.0$3, then use dijets to extract information on both $3.0\!-\!4.0$4 and $3.0\!-\!4.0$5. For cold nuclear matter transport, the nuclear modification factor is written

$3.0\!-\!4.0$6

For spin physics, the proposal highlights single transverse-spin asymmetries

$3.0\!-\!4.0$7

for $3.0\!-\!4.0$8, $3.0\!-\!4.0$9, $pp$0, jets, heavy flavor, and quarkonia, and longitudinal double-spin asymmetries

$pp$1

alongside the spin-decomposition identity

$pp$2

The proposal is explicitly conceptual and uses a “straw man” detector layout. Quantitative performance metrics are described as under study; nevertheless, the intended capabilities are specified qualitatively: sufficiently strong magnetic fields for good forward momentum resolution, sufficient jet resolution in $pp$3 and $pp$4, direct-$pp$5 identification through preshower or charged-particle veto plus EMCal granularity, and significant Drell–Yan yields in $pp$6 GeV $pp$7 with long runs. Hardware and physics synergies with the central sPHENIX barrel are central to the concept, especially for broad rapidity coverage, jet acceptance matching, and combined small-$pp$8 plus sQGP studies.

3. FENIX as an airborne hyperspectral sensor in remote sensing

In remote sensing, FENIX denotes the SPECIM AisaFENIX 1K airborne system mounted on a Cessna 172 and used to generate simulated ground truth for deep-learning classification of VENuS imagery. The study area is a transect in Israel from Beit-Guvrin to Lehavim. The FENIX acquisition on 4 April 2017 covered a 35 km strip with a 1.5 km swath at ground sampling distance 1 m. Spectrally, the sensor spans VNIR $pp$9 nm at 4.5 nm spectral resolution and SWIR $p+A$0 nm at 12 nm spectral resolution; for the study, the data were resampled to 41 bands of 5 nm width covering $p+A$1 nm (Faran et al., 2019).

The key methodological role of FENIX is not merely acquisition but ground-truth simulation. The study avoids manual labeling by unmixing the high-resolution hyperspectral data with a linear mixture model. In canonical notation,

$p+A$2

with abundance constraints $p+A$3; the study uses

$p+A$4

Abundances are estimated with VPGDU, a projected-gradient-descent unmixing method with exact line search and a Spectral Angle Mapper objective,

$p+A$5

The seven predefined endmember classes are Brown Soil, Light Soil, Rock, Tall Tree/Shrub, Dwarf Shrub, Herbaceous, and Dense Shrub/Burned Area.

The simulated ground truth is created by harmonizing FENIX to VENuS both spectrally and spatially. Spectrally, bands are selected in the $p+A$6 nm range to match VENuS, and VENuS band 6 is removed because it duplicates band 5, leaving 11 bands. Spatially, $p+A$7 FENIX pixels at 1 m are aggregated into one VENuS-like 5 m pixel. For each aggregated region, abundances are combined and the label is set by argmax: $p+A$8 The paper notes no explicit SRF convolution and no explicit PSF/MTF modeling.

The downstream classifier is a patch-based CNN for pixel-wise classification. Input patches have size $p+A$9. The architecture consists of 4 convolution layers with $A+A$0 kernels and filter counts $A+A$1, batch normalization, ReLU, dropout at rate $A+A$2, and 3 fully connected layers with a 7-neuron softmax output. Band-wise standardization to zero mean and unit variance is used. Training employs cross-entropy loss,

$A+A$3

with Adam at learning rate $A+A$4, batch size 64, and 200 epochs. Augmentation includes flips, $A+A$5 rotations, and additive Gaussian noise with mean 0 and standard deviation 0.1; class balancing uses 30,000 samples per label per epoch.

The reported quantitative results are limited but explicit. In 6-fold cross-validation on FENIX-simulated data, the average validation accuracy is $A+A$6 and the average test accuracy is $A+A$7. Per-scene test accuracies are 69.40% for Amazya1, 69.97% for Avisure1, 75.08% for Avisure2, 70.96% for Between1, 73.27% for Between2, and 76.71% for Lehavim1. Transfer to real VENuS imagery is zero-shot: the model trained on simulated data is applied directly to an atmospherically corrected VENuS L1 image, and the reported assessment is visual rather than metric-based.

The paper also states several limitations. Radiometric characteristics such as SNR and dynamic range are not explicitly reported for the AisaFENIX in the paper; SRFs are not reported; per-class accuracy, $A+A$8, Cohen’s $A+A$9, and confusion matrices are not reported; and the use of linear unmixing, the $x$0 constraint, and argmax labeling can introduce label noise, especially for mixed pixels, edges, and illumination variation. The significance of FENIX here is therefore as a high-resolution, hyperspectral source for reproducible, spatially dense simulated labels rather than as the object of classification itself.

4. FENIX in programmable data planes and in-network DNN inference

In networking, FENIX is a hybrid in-network ML system that performs feature extraction on programmable switch ASICs and deep neural network inference on FPGAs. The architecture places a Data Engine on a Tofino ASIC and a Model Engine on a Xilinx ZU19EG FPGA integrated directly on the same 22-layer PCB. The design objective is to reconcile low latency, high throughput, and high accuracy under the fundamental mismatch between multi-terabit switch ASICs and FPGA inference rates in the hundreds-of-Gbps regime (Gao et al., 20 Jul 2025).

The Data Engine is built from a Flow Tracker, a probabilistic token-bucket Rate Limiter, and a Buffer Manager. The Flow Tracker maintains a Flow Info Table in SRAM indexed by truncated five-tuple hash and stores hash, backlog packet count $x$1, backlog timestamp $x$2, classification class, and a ring-buffer index $x$3. The Buffer Manager holds the last $x$4 features per flow, with $x$5 in the buffer plus the current feature in metadata, illustrated as a 9-element window $x$6. On a send decision, the selected window slice is serialized into a custom header in the Deparser and mirrored to the FPGA.

The central control mechanism is the probabilistic token bucket. Global flow count $x$7 and packet rate $x$8 are measured over a window $x$9. For flow $\eta = -\ln \tan(\theta/2)$0, $\eta = -\ln \tan(\theta/2)$1 is the elapsed time since last feature transmission and $\eta = -\ln \tan(\theta/2)$2 is the number of packets since that transmission, so

$\eta = -\ln \tan(\theta/2)$3

The token generation rate is

$\eta = -\ln \tan(\theta/2)$4

where $\eta = -\ln \tan(\theta/2)$5 is FPGA-side processing frequency in vectors per second, $\eta = -\ln \tan(\theta/2)$6 is link bandwidth, and $\eta = -\ln \tan(\theta/2)$7 is feature-vector width. The send probability is piecewise linear: $\eta = -\ln \tan(\theta/2)$8 The expected inter-send period is

$\eta = -\ln \tan(\theta/2)$9

The paper states that this construction satisfies equal-rate fairness on average, provides proportional allocation when $\eta$0 differs across flows, and avoids oversubscription by capping bucket capacity to queue length.

The Model Engine implements INT8 DNN inference on a shared systolic array. Supported operators include embedding lookups, fully connected layers, convolution, and recurrent layers, represented generically as

$\eta$1

The evaluated models are a CNN with 3 convolution layers of 64, 128, and 256 filters followed by FC layers of 512 and 256, and an RNN with a single custom RNN cell of 128 units and a dense output layer. Inputs are protocol-agnostic sequences of packet lengths and inter-packet delays; the switch supplies a fixed $\eta$2-packet window per flow. Quantization is offline INT8 using Vitis-AI with per-layer decimal points tuned from activation distributions.

The reported measurements are hardware-specific. Average inference time is approximately $\eta$3; internal on-board transfers are sub-$\eta$4; and external optical-module paths add roughly $\eta$5 when present. On the ZU19EG, the overall CNN uses 38.4% LUT, 33.8% FF, 7.1% BRAM, and 8.1% DSP; the RNN uses 25.6% LUT, 31.2% FF, 6.3% BRAM, and 4.6% DSP; Vector I/O uses 6.0% LUT and 4.8% FF. On Tofino 2, P4 overhead is 12.9% SRAM, 4.4% TCAM, 3.5% bus, and 9 pipeline stages.

Evaluation uses ISCXVPN2016 and USTC-TFC2016. On ISCXVPN2016, macro-$\eta$6 is 0.912 for FENIX-F-RNN and 0.892 for FENIX-P-CNN, compared with 0.870 for FlowLens, 0.863 for BoS, 0.738 for N3IC, 0.658 for NetBeacon, and 0.578 for Leo. On USTC-TFC2016, macro-$\eta$7 is 0.907 for FENIX-P-CNN and 0.901 for FENIX-F-RNN, compared with 0.914 for FlowLens, 0.858 for N3IC, 0.814 for BoS, 0.741 for Leo, and 0.670 for NetBeacon. The hardware tests saturating a 400G-class NIC show no accuracy degradation, while large-scale simulation shows only approximately 13.2% macro-$\eta$8 drop at the largest scale. In this literature, FENIX therefore names a hardware co-design whose defining feature is principled feature metering between line-rate packet processing and low-latency FPGA inference.

5. Fenix as a cosmological hydrodynamical simulation suite

In cosmology, Fenix is a simulation suite rather than an instrument or hardware platform. The cited study analyzes the hydrodynamic run S230D and its matched dark-matter-only counterpart using GADGET-3 in a $\eta$9CDM cosmology with $\theta$0, $\theta$1, $\theta$2, $\theta$3, and $\theta$4. The simulated volume is a periodic box of side 14 Mpc with $\theta$5 particles, mass resolutions $\theta$6 and initial $\theta$7, and maximum Plummer-equivalent gravitational softening 0.35 kpc. The baryonic subgrid model includes metal-dependent radiative cooling, a multiphase interstellar medium, stochastic star formation, chemical enrichment, and supernova feedback, but no AGN feedback (Cataldi et al., 2020).

Haloes are identified with Friends-of-Friends and SUBFIND, with centres refined by the shrinking-sphere method. Hydro and DMO haloes are matched one-to-one from shared initial conditions, and DMO masses are rescaled by the dark-matter fraction $\theta$8, where $\theta$9. The selected $1.2\!-\!3.0$00 sample requires more than 10,000 baryonic particles within the optical radius $1.2\!-\!3.0$01, yielding 38 haloes in the mass range $1.2\!-\!3.0$02.

The study defines virial quantities through

$1.2\!-\!3.0$03

with $1.2\!-\!3.0$04 and $1.2\!-\!3.0$05, using $1.2\!-\!3.0$06 and $1.2\!-\!3.0$07. Inner mass response is quantified at 5%, 10%, and 20% of $1.2\!-\!3.0$08 by

$1.2\!-\!3.0$09

Halo shapes are measured with an iterative ellipsoidal-shell procedure based on the reduced inertia tensor, yielding principal axes $1.2\!-\!3.0$10, axis ratios $1.2\!-\!3.0$11 and $1.2\!-\!3.0$12, and triaxiality

$1.2\!-\!3.0$13

Velocity anisotropy is

$1.2\!-\!3.0$14

Galaxy morphology is characterized by the stellar disc-to-total ratio $1.2\!-\!3.0$15, defined kinematically using the circularity parameter $1.2\!-\!3.0$16, with stars of $1.2\!-\!3.0$17 assigned to the disc.

The principal result is that baryons significantly affect the inner halo, mainly within approximately 20 percent of the virial radius. In Fenix, inner haloes almost always contract in Hydro relative to DMO, unlike EAGLE where many haloes expand in the very centre. Hydro haloes are more spherical, with higher $1.2\!-\!3.0$18 and $1.2\!-\!3.0$19, and more oblate, with lower $1.2\!-\!3.0$20, than their DMO counterparts. The largest deviations occur near $1.2\!-\!3.0$21, while Hydro and DMO shapes converge near $1.2\!-\!3.0$22. The amplitude of the Hydro–DMO shape change is larger in Fenix than in EAGLE, consistent with stronger central baryon concentrations in Fenix.

The paper also identifies a morphology–halo connection. Disc-dominated galaxies preferentially form in DMO haloes that are already more spherical and more oblate within $1.2\!-\!3.0$23. At $1.2\!-\!3.0$24, the Fenix Hydro correlations are reported as Spearman $1.2\!-\!3.0$25 for $1.2\!-\!3.0$26 versus $1.2\!-\!3.0$27 with $1.2\!-\!3.0$28, $1.2\!-\!3.0$29 for $1.2\!-\!3.0$30 versus $1.2\!-\!3.0$31 with $1.2\!-\!3.0$32, and $1.2\!-\!3.0$33 for $1.2\!-\!3.0$34 versus $1.2\!-\!3.0$35 with $1.2\!-\!3.0$36. Velocity anisotropy is also morphology-dependent: inside $1.2\!-\!3.0$37, haloes hosting discier galaxies exhibit higher $1.2\!-\!3.0$38, and baryons reduce $1.2\!-\!3.0$39, making orbits more tangentially biased, though this reduction is weaker for disc-dominated systems. The paper interprets these trends as evidence for a cosmological connection between final galaxy morphology and intrinsic dark-halo structure, strengthened by disc growth.

6. Conceptual boundaries, recurring methodological motifs, and common confusions

The four usages of FENIX occupy different epistemic roles. In the RHIC literature, FENIX is a detector concept for future measurements; in remote sensing, it is an airborne sensor whose data are transformed into simulated labels; in networking, it is a complete hardware–software inference platform; in cosmology, it is a simulation suite used to compare Hydro and DMO realizations. Any attempt to merge them into a single technical lineage would be unsupported by the cited papers (Seto, 2012, Faran et al., 2019, Gao et al., 20 Jul 2025, Cataldi et al., 2020).

At the same time, a limited structural comparison is possible. Each usage couples a data-acquisition or data-generation mechanism to a downstream inference or interpretation layer. In fsPHENIX, tracking, calorimetry, PID, and magnetic-field design are organized around observables such as $1.2\!-\!3.0$40, dijets, Drell–Yan, and spin asymmetries. In the VENuS classification workflow, hyperspectral acquisition feeds linear unmixing, then CNN training. In the programmable-switch system, line-rate feature extraction feeds a quantized CNN or RNN through a probabilistic token bucket. In the cosmological suite, hydrodynamical and DMO realizations feed comparative measurements of contraction, shape, and anisotropy. This suggests a common pattern of staged measurement pipelines, but not a common underlying technology.

The most common misconception is lexical rather than scientific: that “FENIX” names one project spanning detector physics, hyperspectral sensing, programmable networks, and galaxy formation. The cited literature does not support that reading. A second potential confusion concerns the PHENIX/sPHENIX usage: what is sometimes called “FENIX” there is specifically the forward sPHENIX spectrometer, or fsPHENIX, rather than the central detector as a whole. A third concerns the remote-sensing paper: FENIX is the source of high-resolution hyperspectral data used to simulate ground truth, whereas the transfer target is VENuS. A fourth concerns the cosmology paper: Fenix is the name of the simulation suite, not a derived halo-shape metric.

Taken together, these literatures make FENIX a notable example of cross-domain name reuse in contemporary technical writing. The term acquires meaning only through its local disciplinary embedding: rapidity coverage and small-$1.2\!-\!3.0$41 QCD at RHIC, hyperspectral unmixing and CNN transfer in remote sensing, probabilistic feature metering for FPGA-assisted in-network inference, or baryon-driven reshaping of dark-matter haloes in cosmological simulation.