KRATOS Suite: Multi-Domain Research Systems

• KRATOS Suite is a polysemous term covering domain-specific systems, including GPU-first astrophysical frameworks, N-body simulation libraries, FPGA benchmarks, and IoT platforms.
• In astrophysics, it offers high-performance, multiphysics simulation modules for hydrodynamics, radiative transfer, and reactive flows, as well as detailed N-body analyses of galactic dynamics.
• Beyond astrophysics, KRATOS supports circuit-level FPGA benchmarking for sparse, low-precision DNNs, rapid LPWAN prototyping, and smart-home access control through integrated hardware-software stacks.

Searching arXiv for papers on "KRATOS" across the domains represented in the provided data. KRATOS Suite is a polysemous research designation applied to several technically distinct systems across astrophysical simulation, galactic dynamics, FPGA benchmarking, LPWAN prototyping, and smart-home access control. In current scholarly usage, the name most prominently denotes a family of GPU-first heterogeneous astrophysical frameworks and associated modules, including hydrodynamics, ray tracing, thermochemistry, and polarized Monte Carlo radiative transfer, as well as an open suite of pure N-body simulations for interpreting Large Magellanic Cloud kinematics (Wang, 4 Jan 2025, Wang, 7 Apr 2025, Yang et al., 1 Dec 2025, Jiménez-Arranz et al., 2024, Schölch et al., 2 Aug 2025). The same name is also used for a circuit-level FPGA benchmark for unrolled deep neural networks (Dai et al., 2024), an open-source LoRa hardware-software platform (Piyare et al., 2018), and a multi-user, multi-device-aware access control system for smart homes (Sikder et al., 2019). The term therefore does not denote a single unified software product across all fields; rather, it names several domain-specific suites whose internal architectures, performance objectives, and scientific roles are unrelated except by nomenclature.

1. Astrophysical heterogeneous computing framework

In astrophysics, Kratos is a modern, GPU-optimized, general-purpose simulation framework for astrophysical fluid and multiphysics problems designed explicitly for heterogeneous architectures (Wang, 4 Jan 2025). It targets both NVIDIA (CUDA) and AMD (HIP) GPUs and retains development-time and testing compatibility on CPUs via HIP-CPU, while being written in C++17 with a thin device abstraction layer that exposes device initialization and finalization, device dispatch, shared-memory directives, memory allocation and deallocation, host-device copies, kernel launches, streams, and events (Wang, 4 Jan 2025). The framework is described as “module-defined”: modules are decoupled from mesh structures and can be orchestrated in arbitrary sequences via a module container or “task lists,” allowing pure radiative transfer, hydrodynamics, or more complex multiphysics assemblies (Wang, 4 Jan 2025).

The core infrastructure combines a device abstraction layer, a stream-aware MPI communication model, and a mesh management system serving as the basis for physical modules (Wang, 4 Jan 2025). The multiprocessing backend uses MPI-3 by default and queues nonblocking communication onto device streams so that dependencies and ordering are respected without forced host-side synchronization, explicitly aiming to overlap communication and computation (Wang, 4 Jan 2025). Mesh management is based on structured grids organized as $2^d$ trees, with octrees in 3D, quadtrees in 2D, and binary trees in 1D; refinement replaces a coarse block by $2^d$ child blocks at half size and double resolution, with touching blocks differing by at most one level (Wang, 4 Jan 2025). Partitioning may use lexicographic row-filling or Hilbert space-filling curves, with optional weighting to account for device heterogeneity (Wang, 4 Jan 2025).

The hydrodynamics module provides the foundational Godunov solver. It evolves the Euler equations in conservative form with state vector $U = [\rho, \rho v, E]^T$ and total energy $E = p/(\gamma - 1) + (1/2)\rho ||v||^2$, using finite-volume discretization, PLM reconstruction, an HLLC Riemann solver, and second-order Runge–Kutta integration (Wang, 4 Jan 2025). Cross-level AMR interfaces use volume-weighted restriction, PLM prolongation with minmod slopes, and flux reconciliation so that conservation across levels is enforced (Wang, 4 Jan 2025). Mixed precision is central to the design: reconstruction and HLLC flux estimates are computed in single precision, while conservative variables and their updates are stored and accumulated in double precision, preserving mass, momentum, and energy to machine precision of FP64 (Wang, 4 Jan 2025).

Validation is reported through shock tubes, double Mach reflection, Kelvin–Helmholtz instability, and Rayleigh–Taylor instability (Wang, 4 Jan 2025). Throughput figures are given in the range $10^8$–$10^9$ cells s$^{-1}$ for hydrodynamic benchmarks, with mixed precision typically running at about $70$–$80\%$ of single-precision speed on consumer GPUs and about $2$–$2^d$0 faster than full double precision (Wang, 4 Jan 2025). This combination of GPU-near programming style, stream-aware communication, and module isolation defines the architectural core from which later Kratos astrophysical modules extend.

2. Radiation, reacting flow, thermochemistry, and polarized radiative transfer

A second major layer of the astrophysical Kratos suite adds thermochemistry, ray tracing, radiation-matter interaction, and polarized radiative transfer (Wang, 7 Apr 2025, Yang et al., 1 Dec 2025). The reacting-flow and thermochemistry extension integrates hydrodynamics with cell-local stiff ODE solves and direct ray tracing on structured meshes (Wang, 7 Apr 2025). It introduces a stoichiometry-compatible reconstruction scheme for chemical species advection in which reconstruction increments are projected into the stoichiometric nullspace so that elemental conservation is maintained without per-face matrix inversions (Wang, 7 Apr 2025). The governing elemental constraint is written as $2^d$1, and the projection operator is precomputed via an SVD of the element composition matrix (Wang, 7 Apr 2025). Species fluxes are then updated conservatively, with the method retaining high-order accuracy while avoiding the $2^d$2 cost of repeated inversions (Wang, 7 Apr 2025).

Thermochemistry is operator-split from hydrodynamics and solved semi-implicitly with adaptive substeps (Wang, 7 Apr 2025). The framework assembles analytically dominated Jacobians for species densities and internal energy, includes temperature-dependent rate derivatives, and employs a custom Crout LU decomposition with row pivoting tailored for batches of many small matrices on SIMT hardware (Wang, 7 Apr 2025). Radiation is treated through a ray-marching method on structured meshes that computes exact cell intersection lengths and photon attenuation, using

$2^d$3

for beamed radiation with absorption and emission dominant (Wang, 7 Apr 2025). Verification spans chemical advection, combustion, Strömgren spheres, and detonation dynamics, with results reported as closely matching semi-analytic solutions, Cantera, and the Shock and Detonation Toolbox (Wang, 7 Apr 2025).

Kratos-polrad extends this framework into polarized Monte Carlo radiative transfer (Yang et al., 1 Dec 2025). It is described as a novel GPU-accelerated Monte Carlo Radiative Transfer code built upon the heterogeneous computing framework of Kratos and designed for self-consistent and efficient polarization calculations (Yang et al., 1 Dec 2025). The code uses KRATOS mesh APIs to access cell-centered fields such as extinction mean free paths $2^d$4 and $2^d$5, camera geometry, and optional magnetic-field direction per cell, while packet propagation relies on the existing KRATOS ray-tracing infrastructure to step through cells, compute optical depths, and locate interaction points (Yang et al., 1 Dec 2025). The problem addressed is efficient, self-consistent polarized Monte Carlo radiative transfer in 3D geometries with multiple scattering and dichroic extinction, which had previously been prohibitive at high photon counts on CPUs (Yang et al., 1 Dec 2025).

A defining feature is full Stokes tracking throughout propagation, with $2^d$6 transported under extinction, emission, and scattering (Yang et al., 1 Dec 2025). The vector radiative transfer equation is expressed as

$2^d$7

with $2^d$8 and $2^d$9 (Yang et al., 1 Dec 2025). For dichroic media, the extinction matrix in optical-depth form contains block couplings between $U = [\rho, \rho v, E]^T$0 and $U = [\rho, \rho v, E]^T$1 through coefficients $U = [\rho, \rho v, E]^T$2, $U = [\rho, \rho v, E]^T$3, and $U = [\rho, \rho v, E]^T$4 (Yang et al., 1 Dec 2025). Scattering uses a Mueller matrix $U = [\rho, \rho v, E]^T$5 in the local grain frame, after which $U = [\rho, \rho v, E]^T$6 and $U = [\rho, \rho v, E]^T$7 are rotated into the new polarization basis with

$U = [\rho, \rho v, E]^T$8

and $U = [\rho, \rho v, E]^T$9 defined through $E = p/(\gamma - 1) + (1/2)\rho ||v||^2$0 and $E = p/(\gamma - 1) + (1/2)\rho ||v||^2$1 entries (Yang et al., 1 Dec 2025).

The grain-frame transforms use quaternion algebra, with a pure-quaternion sandwich product $E = p/(\gamma - 1) + (1/2)\rho ||v||^2$2 and a specific convention for the quaternion components when grains are perfectly aligned to the local magnetic field described by $E = p/(\gamma - 1) + (1/2)\rho ||v||^2$3 (Yang et al., 1 Dec 2025). A further innovation is an analytic per-cell solution for optically thick segments. Rather than using naive linearized updates, Kratos-polrad solves the coupled Stokes evolution analytically over each cell, yielding explicit expressions for $E = p/(\gamma - 1) + (1/2)\rho ||v||^2$4, $E = p/(\gamma - 1) + (1/2)\rho ||v||^2$5, $E = p/(\gamma - 1) + (1/2)\rho ||v||^2$6, and $E = p/(\gamma - 1) + (1/2)\rho ||v||^2$7 involving $E = p/(\gamma - 1) + (1/2)\rho ||v||^2$8 multiplied by hyperbolic or trigonometric functions of $E = p/(\gamma - 1) + (1/2)\rho ||v||^2$9 and $10^8$0 (Yang et al., 1 Dec 2025). This is presented as the mechanism that stabilizes calculations in optically thick cells.

The imaging workflow is two-step polarimetry imaging (Yang et al., 1 Dec 2025). In the first step, Monte Carlo packet propagation continuously accumulates the scattering source function

$10^8$1

with accumulation along path segments rather than only at discrete scattering sites (Yang et al., 1 Dec 2025). In the second step, deterministic line-of-sight ray integration uses the same per-cell analytic Stokes solver to synthesize images, polarization maps, and SEDs for one or multiple cameras from the same Monte Carlo run (Yang et al., 1 Dec 2025). The output includes $10^8$2 images, $10^8$3, position angles, and SEDs (Yang et al., 1 Dec 2025).

Validation covers self-scattering polarization in inclined disks, comparison against RADMC-3D for a dust cylinder with thermal emission, twisted magnetic fields in a slab, dichroic extinction in a uniformly aligned slab, and scattering of central starlight by aligned grains in a uniform sphere (Yang et al., 1 Dec 2025). Quantitatively, the code is reported to match the analytical polarization fraction

$10^8$4

up to $10^8$5, whereas RADMC-3D deviates at $10^8$6 (Yang et al., 1 Dec 2025). GPU parallelism is organized with one warp or wavefront per packet, shared-memory reductions for source-function accumulation, and lightweight deterministic imaging kernels that reuse grid traversal routines (Yang et al., 1 Dec 2025). Benchmark figures include $10^8$7 packet sampling of the scattering source function in about $10^8$8 s on a single RTX 5090 and an overall speedup of about $10^8$9 relative to CPU codes for polarized MCRT (Yang et al., 1 Dec 2025).

A plausible implication is that, within astrophysics, “KRATOS Suite” now denotes not only a hydrodynamics framework but a modular simulation environment spanning conservative fluid dynamics, stiff microphysics, ray-based radiation transport, and full-Stokes polarized radiative transfer on the same heterogeneous infrastructure.

3. KRATOS N-body simulations for the Magellanic system

A separate astrophysical use of the name refers to KRATOS as an open-access suite of 28 high-resolution, pure N-body simulations designed to interpret the morphologies and kinematics of LMC-like stellar discs under tidal interaction with an SMC-mass companion and, in some configurations, a MW-mass host (Jiménez-Arranz et al., 2024). In expanded form, the acronym is given as “Kinematic Reconstruction of the mAgellanic sysTem within the OCRE Scenario” (Jiménez-Arranz et al., 2024). The suite is organized into 11 sets, each containing up to three configurations: an isolated LMC-like galaxy, an LMC-like galaxy interacting with an SMC-mass galaxy, and a configuration in which both SMC-mass and MW-mass galaxies may interact with the LMC-like system (Jiménez-Arranz et al., 2024).

The simulations were run with the ART pure N-body AMR gravity solver in a periodic box of $10^9$0 Mpc$10^9$1, with spatial resolution $10^9$2 pc, temporal resolution $10^9$3 yr, and minimum particle mass $10^9$4 for the LMC-like and SMC-mass components (Jiménez-Arranz et al., 2024). The total integration length is $10^9$5 Gyr (Jiménez-Arranz et al., 2024). The fiducial LMC-like galaxy consists of a live halo and exponential stellar disc with scale length $10^9$6 kpc, scale height $10^9$7 kpc, truncation radius $10^9$8 kpc, Toomre $10^9$9, stellar mass $^{-1}$0, and an NFW halo with concentration $^{-1}$1 and total dark-matter mass $^{-1}$2 (Jiménez-Arranz et al., 2024). The SMC-mass galaxy is represented by a live NFW halo of concentration $^{-1}$3 and dark-matter mass $^{-1}$4, with a post hoc most-bound “stellar” subset reaching $^{-1}$5 (Jiménez-Arranz et al., 2024). The MW-mass system is a live NFW halo with concentration $^{-1}$6 and mass $^{-1}$7, without an MW disc or bulge (Jiménez-Arranz et al., 2024).

The suite is explicitly intended for first-order interpretation of observed LMC structures and velocity fields, including bars, spirals, warps, disc thickening, and interaction-driven asymmetries (Jiménez-Arranz et al., 2024). Analysis procedures include centering by the baryonic center of mass or the KDE-derived bar density center, alignment of the disc plane with the stellar angular momentum vector, and bar characterization using the Dehnen method (Jiménez-Arranz et al., 2024). Velocity fields are constructed after subtracting the LMC systemic motion, and cylindrical components $^{-1}$8, $^{-1}$9, and $70$0 are measured in the aligned disc frame (Jiménez-Arranz et al., 2024). Bar detection relies on the relative $70$1 Fourier amplitude, with a barred state defined by

$70$2

the same criterion later reused in follow-on analyses (Jiménez-Arranz et al., 2024).

The key scientific results are morphological and dynamical. Across the 28 simulations, 17 develop bars and 11 do not show a clear bar at any epoch (Jiménez-Arranz et al., 2024). Tidal interactions are reported to boost the disc scale height, both destroy and create bars, and naturally explain off-center stellar bars (Jiménez-Arranz et al., 2024). Off-centeredness of the bar is quantified as

$70$3

with typical peaks of about $70$4–$70$5 kpc about $70$6 Gyr after the first pericenter (Jiménez-Arranz et al., 2024). Long-lived bars are described as robust: their bar length $70$7 and pattern speed $70$8 are not appreciably altered by SMC pericenters, with pattern speeds spanning about $70$9–$80\%$0 km s$80\%$1 kpc$80\%$2 and interaction-induced differences typically only a few km s$80\%$3 kpc$80\%$4 (Jiménez-Arranz et al., 2024). The same suite is made publicly available through a Dataverse repository with DOI 10.34810/data1156 (Jiménez-Arranz et al., 2024).

4. Kinematic asymmetries in the LMC and the role of KRATOS

The N-body KRATOS suite is used directly in the study of asymmetric velocity patterns in the LMC revealed by Gaia DR3 (Schölch et al., 2 Aug 2025). In that context, KRATOS is described as a large suite of high-resolution, collisionless N-body simulations designed specifically to interpret the stellar kinematics and morphology of LMC-like discs in isolation and in interaction with an SMC-mass companion and a MW-mass host (Schölch et al., 2 Aug 2025). The paper analyzes two particular models, K4 and K6, with K6 selected for its morphological similarity to the LMC, including an off-centered bar and a single dominant spiral arm (Schölch et al., 2 Aug 2025). The resolution figures restate the underlying suite parameters: total evolution time $80\%$5 Ga, spatial resolution $80\%$6 pc, temporal resolution $80\%$7 yr, and minimum particle mass $80\%$8 (Schölch et al., 2 Aug 2025).

The methodological contribution is a new procedure to quantify asymmetry in the bar and outer disc (Schölch et al., 2 Aug 2025). For the bar, the simulation or Gaia map is deprojected into an in-plane Cartesian system, a bar contour is defined where the local density exceeds the Gaussian-smoothed background by at least $80\%$9, and the medians of the radial velocity and residual tangential velocity are measured within that contour (Schölch et al., 2 Aug 2025). The in-plane velocities satisfy

$2$0

and the residual tangential velocity is

$2$1

(Schölch et al., 2 Aug 2025). Significant deviations of the bar medians from zero are interpreted as evidence that the bar quadrupole is asymmetric (Schölch et al., 2 Aug 2025). For the outer disc, the analysis masks the inner region, divides the remaining disc into 20 azimuthal wedges of $2$2, computes wedge medians, subtracts opposite wedges, and defines asymmetry measures through the extrema

$2$3

(Schölch et al., 2 Aug 2025).

The study also transforms KRATOS simulations into Gaia mock catalogues (Schölch et al., 2 Aug 2025). The mock pipeline projects particles into the heliocentric frame at the LMC sky position, adopts Red Clump stellar parameters $2$4 and $2$5, uses a 3D extinction model consistent with Gaia Object Generator usage, and applies PyGaia error models for DR3, DR4, and DR5 (Schölch et al., 2 Aug 2025). Proper-motion to tangential-velocity conversion is taken as

$2$6

(Schölch et al., 2 Aug 2025). These mocks are used to test whether the asymmetries persist under realistic observational errors and selection functions (Schölch et al., 2 Aug 2025).

The conclusions are specific. In the K6 interacting run, the bar asymmetry peaks shortly after the second LMC–SMC pericentre at $2$7 Ga, and the Gaia DR3 bar asymmetry levels pass through those peaks, supporting a dynamical origin from recent interaction (Schölch et al., 2 Aug 2025). By contrast, the strong outer-disc asymmetry seen in the Gaia DR3 optimal sample is traced to a classification bias in the neural-network classifier of the optimal sample rather than to genuine LMC–SMC dynamics (Schölch et al., 2 Aug 2025). The paper recommends using K6 at $2$8 Ga, also labelled $2$9 in KRATOS, as the best present-day proxy for the Gaia DR3 asymmetries (Schölch et al., 2 Aug 2025). This establishes KRATOS not only as a static simulation repository but as a dynamical interpretive framework for time-correlated comparison between N-body interaction histories and astrometric surveys.

5. FPGA benchmark suite for unrolled DNNs

Outside astrophysics, “Kratos” denotes a circuit-level FPGA benchmark suite purpose-built for unrolled deep neural network layers that exploit fine-grained unstructured sparsity and mixed arithmetic precision (Dai et al., 2024). Its scope is completely different from the astrophysical software, focusing on DNN kernels whose weights are embedded into LUT logic so that bit-level and parameter-level ineffectual operations can be skipped (Dai et al., 2024). The goals are to quantify area, delay, and frequency characteristics of unrolled DNN primitives under varying sparsity and precision, expose scaling limits of partial versus full unrolling, and enable FPGA architecture exploration tailored to sparse, low-precision unrolled DNNs (Dai et al., 2024).

The benchmark contains eight kernels spanning GEMM and convolution. These include GEMM via multiply-adder trees and weight-stationary systolic arrays, as well as 1-D and 2-D convolution via multiply-adder trees, with pixelwise, row-parallel, and fully-unrolled input-unrolling factors depending on the kernel (Dai et al., 2024). Sparsity is defined as the percentage of zero elements in the weight tensor, with exact numbers of nonzeros randomly shuffled to emulate unstructured sparsity (Dai et al., 2024). Precision experiments use 1-, 2-, 4-, and 8-bit integer data types, though any integer bitwidth is supported (Dai et al., 2024). The methodology targets an Intel Arria 10 GX 1150 and uses Intel Quartus Prime 22.3, Modelsim, the open-source VTR flow with Yosys, Odin II, and VPR, and COFFE 2.0 for architecture exploration on a Stratix-IV-like 40 nm fabric (Dai et al., 2024).

The benchmark’s main quantitative findings are circuit-level. Fully-unrolled designs reach the Arria 10 device maximum, with unrestricted $2^d$00 GHz and restricted $2^d$01 MHz, whereas row-parallel and pixelwise designs commonly achieve about $2^d$02–$2^d$03 MHz (Dai et al., 2024). For tree-based implementations, normalized ALM utilization tracks density closely, yielding near-ideal linear area reduction with sparsity; by contrast, systolic GEMM retains structural registers and control, so at $2^d$04 sparsity ALM usage falls by only $2^d$05 for the small case and $2^d$06 for the large case (Dai et al., 2024). In one illustrative example, a $2^d$07 matrix multiply on Arria 10 uses $2^d$08 of the FPGA under naive full unrolling, while specialization plus pruning plus quantization reduces area by roughly $2^d$09 in the text and up to $2^d$10 in a figure caption, down to about $2^d$11 of the FPGA for 4096 effective FLOPs (Dai et al., 2024).

The architecture study varies LUT size $2^d$12 and shows that reducing LUT size from $2^d$13 to $2^d$14 yields about $2^d$15 silicon-area reduction across representative kernels with modest frequency impact (Dai et al., 2024). Example figures include conv2d-PW-S shrinking from $2^d$16 mm$2^d$17 to $2^d$18 mm$2^d$19 and gemmt-RP-S from $2^d$20 mm$2^d$21 to $2^d$22 mm$2^d$23 (Dai et al., 2024). The normalized area-delay product generally favors $2^d$24 (Dai et al., 2024). Source code is released at github.com/abdelfattah-lab/Kratos-benchmark (Dai et al., 2024). In this domain, therefore, KRATOS Suite designates a reproducible, parameterized benchmark environment for co-designing FPGA fabrics and unrolled sparse/mixed-precision DNN circuits rather than a runtime simulation system.

6. LPWAN platform and smart-home access-control suite

Two additional research systems use the KRATOS name in embedded and IoT contexts. The first is an open-source hardware-software platform for rapid research in LPWANs, centered on a low-power multi-sensor LoRa node with an ultra-low-power wake-up receiver and a ContikiOS-based networking stack (Piyare et al., 2018). Its hardware uses a Texas Instruments MSP430FR5969 MCU and a Semtech SX1276 transceiver supporting LoRa, (G)FSK, and OOK (Piyare et al., 2018). The wake-up receiver is a passive OOK demodulator front-end tuned for 868 MHz, with a PIC12LF1552 for address decoding, standby power of $2^d$25W, decoding power of $2^d$26W, sensitivity $2^d$27 dBm, and maximum bit rate $2^d$28 kbps (Piyare et al., 2018). The power-management unit is based on a TI BQ25570 with MPPT and about $2^d$29 efficiency, supporting recharging of supercapacitors and a Li-Ion battery (Piyare et al., 2018). Reported system-level consumption spans from $2^d$30W in deep sleep with wake-up listening to $2^d$31 mW during LoRa transmission at $2^d$32 dBm (Piyare et al., 2018).

On the software side, the platform adds a ContikiOS port for the MSP430FR5xx family and a radio abstraction layer adapting Semtech’s SX1276 driver to Contiki’s API (Piyare et al., 2018). Through this layer, LoRa parameter settings such as bandwidth, spreading factor, coding rate, frequency, and power can be configured, while unmodified Contiki upper-layer protocols such as Rime and uIPv6 become available over the LoRa PHY (Piyare et al., 2018). In a network coverage experiment, with transmit power $2^d$33 dBm, bandwidth $2^d$34 kHz, spreading factor $2^d$35, coding rate $2^d$36, and 16-byte packets every 10 s, the platform achieved more than $2^d$37 packet reception at 600 m in a non-line-of-sight environment, with RSSI above $2^d$38 dBm and SNR between 32 and $2^d$39 (Piyare et al., 2018). This use of KRATOS emphasizes open hardware, low-power duty cycling through a wake-up receiver, and an immediately programmable embedded network stack.

The second IoT use is “KRATOS: Multi-User Multi-Device-Aware Access Control System for the Smart Home” (Sikder et al., 2019). Here, KRATOS is a policy-centric suite with four components: a User Interaction Module, Back-end Module, Policy Manager, and Policy Execution Module (Sikder et al., 2019). It formalizes policy clauses as quintuples $2^d$40 over assigner, assignee, device, context, and action class, where the action is either demand or restrict (Sikder et al., 2019). These clauses are translated into ABAC-like rules, and decisions are made over a negotiated final policy set $2^d$41 through a binary allow/deny function $2^d$42 (Sikder et al., 2019).

The technical novelty lies in conflict detection and negotiation (Sikder et al., 2019). Clauses interfere if they concern the same user and device. Hard and soft conflicts are distinguished according to overlap properties in their condition sets, and further classified into Hard Priority Conflict, Soft Priority Conflict, Hard Competition Conflict, Soft Competition Conflict, and Restriction Conflict based on the users’ priority classes (Sikder et al., 2019). The negotiation function $2^d$43 resolves conflicts by priority selection, condition-set intersection, majority vote on binary devices, or arbitration on numeric ranges (Sikder et al., 2019). Implementation is integrated into the Samsung SmartThings ecosystem via customized Groovy SmartApps (Sikder et al., 2019).

Evaluation uses 43 smart-home users grouped into 14 groups, along with an extended dataset of 309 policies including 213 demand conflicts and 24 restriction policies, and a study dataset of 219 policies including 146 demand conflicts and 33 restriction policies (Sikder et al., 2019). Across six scenarios, KRATOS reports 100% conflict resolution, with average negotiation times ranging from 0.2 s to 1.2 s depending on scenario complexity (Sikder et al., 2019). Detection success across five access-control threats is reported as 100%, with average detection times between 0.25 s and 0.47 s and notification times between 0.40 s and 0.60 s (Sikder et al., 2019). Policy execution overhead raises app latency from 1.30 s with no policy to at most 1.92 s with time and value constraints, with CPU usage up to 2.5% and RAM up to 2.82% (Sikder et al., 2019). In this field, KRATOS Suite refers to a complete front-end, policy-analysis, and enforcement system for negotiated smart-home access control.

7. Nomenclature, distinctions, and scope

Because the name appears in multiple unrelated literatures, technical precision requires domain disambiguation. The astrophysical Kratos framework is a GPU-first heterogeneous simulation system with a module container, stream-aware MPI, structured-mesh management, and physics modules including hydrodynamics, reacting flow, ray tracing, thermochemistry, and polarized radiative transfer (Wang, 4 Jan 2025, Wang, 7 Apr 2025, Yang et al., 1 Dec 2025). The KRATOS N-body suite, although also astrophysical, is a distinct line of work: a public library of 28 collisionless simulations for LMC-like discs interacting with SMC- and MW-mass systems, used to interpret morphology and velocity maps rather than to provide a general multiphysics runtime (Jiménez-Arranz et al., 2024, Schölch et al., 2 Aug 2025). The FPGA Kratos benchmark is a circuit-level exploration suite for sparse and mixed-precision unrolled DNN kernels on FPGAs (Dai et al., 2024). The LPWAN KRATOS platform is an open-source LoRa mote and ContikiOS port (Piyare et al., 2018). The smart-home KRATOS system is a policy negotiation and enforcement architecture for multi-user device access (Sikder et al., 2019).

A common misconception would be to treat all of these as components of one overarching software ecosystem. The published record does not support that interpretation. The shared name masks separate projects with different codebases, hardware targets, mathematical formalisms, and scientific goals. What unifies them is only the recurrent label “KRATOS,” not a common architecture or research program across disciplines.

Within individual domains, however, the word “suite” is used in a stronger sense. In astrophysics, it denotes either a modular multiphysics framework or a systematically varied library of simulations (Wang, 4 Jan 2025, Jiménez-Arranz et al., 2024). In FPGA research, it denotes a parameterized benchmark suite with automated CAD-flow generation (Dai et al., 2024). In LPWANs and smart homes, it denotes integrated hardware-software or policy-enforcement stacks (Piyare et al., 2018, Sikder et al., 2019). This suggests that the semantic core of “KRATOS Suite” is methodological completeness inside a domain: a combination of infrastructural components, reproducible workflows, and evaluative apparatus assembled around a specialized research problem.