KineticSim: A Cross-Domain Simulation Paradigm

• KineticSim is a versatile kinetic simulation framework that represents unresolved physics through kinetic state evolution via transport, collision, relaxation, and event rules.
• It employs methodologies ranging from particle-in-cell and gyrokinetic formulations to gas-kinetic CFD, kinetic Monte Carlo, and GPU-accelerated market execution to achieve high-fidelity, multi-regime insights.
• By bridging kinetic and macroscopic approaches, KineticSim enables regime transitions and efficient shared-state simulations, underpinning applications in plasma turbulence, biochemical kinetics, and financial markets.

In the literature summarized here, KineticSim does not denote a single universally standardized simulator. The name is used explicitly for a custom CUDA execution engine for large ensembles of financial limit-order-book markets, and it is also used descriptively for a broader family of simulation frameworks whose governing objects are particles, distribution functions, relaxation operators, or event-driven kinetic updates rather than exclusively macroscopic closures. This suggests that KineticSim is best understood as a cross-domain kinetic-simulation paradigm spanning plasma turbulence, gyrokinetics, gas-kinetic CFD, reaction-diffusion particle systems, kinetic Monte Carlo, and GPU-resident multi-agent simulation (Jayakody et al., 19 Jun 2026, Groselj et al., 2017, Righi, 2013).

1. Scope and methodological identity

Across the cited work, KineticSim-like frameworks share a common architectural choice: they represent unresolved physics through a kinetic state and evolve that state by transport, collision, relaxation, or event rules. In one branch this is a fully kinetic 3D particle-in-cell simulation of subion solar-wind turbulence in OSIRIS (Groselj et al., 2017); in another it is a particle-in-cell $\delta f$ gyrokinetic formulation in GTC for the kinetic infernal mode (Li et al., 2024); in another it is a BGK-type or Anderson–Witting RTA solver for compressible or relativistic flow (Righi, 2013, Ambrus et al., 2022). Particle-based biochemical and Monte Carlo variants replace continuum closures by proximity rules, stochastic event times, or asymptotic-preserving jump–diffusion updates (Kearney et al., 2023, Kearney et al., 31 Jul 2025, Lappi et al., 23 Sep 2025).

A common misconception is that “kinetic” implies a single level of fidelity. The corpus instead ranges from first-principles fully kinetic PIC to global linear gyrokinetics, Legendre-harmonic Vlasov–Fokker–Planck–Boltzmann transport, gas-kinetic flux construction, event-driven particle reaction rules, and persistent shared-memory agent reductions. This suggests that the decisive feature is not a unique discretization, but the use of a kinetic state as the primary computational object (Mijin et al., 2020, Jayakody et al., 19 Jun 2026).

2. Plasma and fusion realizations

In space-plasma turbulence, a fully kinetic KineticSim realization is given by the 3D electromagnetic PIC study of kinetic Alfvén turbulence in a triply periodic box with $L_\perp=16.97\,d_i$, $L_z=42.43\,d_i$, $\mathbf B_0=B_0\hat{\mathbf e}_z$, Maxwellian ions and electrons, $\beta_i=0.5$, $m_i/m_e=64$, and initial counterpropagating oblique Alfvén waves. The simulation shows subion-range magnetic steepening consistent with the frequently reported solar-wind slope near $E(k_\perp)\propto k_\perp^{-2.8}$, density–magnetic equipartition in the KAW normalization, local anisotropy with $k_\parallel<k_\perp$, broad consistency with $k_\parallel\propto k_\perp^{1/3}$, and a nonlinearity parameter $\chi$ of order unity, supporting a critically balanced KAW cascade (Groselj et al., 2017). The paper is explicit that the subion-range scale separation is limited and that alternative physics, including whistler turbulence or reconnection-mediated cascades, may become relevant deeper in the subion range.

In tokamak microstability, the same kinetic emphasis appears in global linear gyrokinetic simulations of the kinetic infernal mode. Using the electromagnetic gyrokinetic model in GTC, the study evolves gyrocenter distribution functions in 5D phase space with parallel Ampère’s law and the gyrokinetic Poisson equation, together with a fluid-kinetic hybrid electron model. The principal result is a smooth KBM-to-KIM transition as magnetic shear weakens: for moderate or strong shear the instability is KBM, whereas for $L_\perp=16.97\,d_i$0 it becomes KIM, with growth rate and real frequency varying continuously and nonmonotonically with a turning point around $L_\perp=16.97\,d_i$1. The organizing parameter is $L_\perp=16.97\,d_i$2, with KBM-like behavior when $L_\perp=16.97\,d_i$3 and KIM-like behavior when $L_\perp=16.97\,d_i$4, implying that the spacing of rational surfaces relative to the mode envelope controls the structural transition (Li et al., 2024).

A further plasma realization is SOL-KiT, a fully implicit 1D code for kinetic simulation of parallel electron transport in the tokamak scrape-off layer. It supports kinetic or fluid electrons, fluid or stationary ions, and diffusive neutrals; expands the electron distribution in an arbitrary-degree Legendre basis; includes electron-ion and electron-atom collisions; introduces a conservative treatment of inelastic and superelastic collisions on nonuniform velocity grids; and gives a full Legendre-form treatment of the logical sheath boundary condition. Verification includes Spitzer–Härm heat-flux recovery, conservative neutral/electron benchmarks with errors near roundoff in integrated particle and energy balances, ion acoustic wave propagation, and sheath-boundary tests (Mijin et al., 2020). Taken together, these plasma examples show that KineticSim-like practice spans first-principles turbulence, linear stability, and non-local edge transport.

3. Gas-kinetic and lattice-kinetic flow simulation

In compressible CFD, a KineticSim interpretation is provided by the gas-kinetic turbulence scheme based on the BGK equation

$L_\perp=16.97\,d_i$5

Here the numerical flux is constructed from moments of the distribution function rather than from a Riemann solver, and turbulence enters through a turbulent relaxation time. The paper distinguishes the trivial analogue $L_\perp=16.97\,d_i$6 from a turbulence-dependent relaxation formulation such as

$L_\perp=16.97\,d_i$7

and introduces a degree of rarefaction

$L_\perp=16.97\,d_i$8

At low rarefaction the method collapses to something very close to a standard Navier–Stokes flux, whereas at higher rarefaction kinetic corrections become active. Numerical comparisons against a Roe-based finite-volume Navier–Stokes solver show differences in shock position, downstream pressure recovery, and separation behavior even when the turbulence closure is the same (Righi, 2013).

A relativistic extension appears in the fast kinetic simulator for relativistic matter, which solves the Anderson–Witting relaxation-time approximation

$L_\perp=16.97\,d_i$9

with a high-order off-lattice relativistic lattice Boltzmann discretization. The method is designed to cover both $L_z=42.43\,d_i$0 fluid regimes and $L_z=42.43\,d_i$1 rarefied-gas regimes, and both $L_z=42.43\,d_i$2 ultra-relativistic and $L_z=42.43\,d_i$3 mildly/non-relativistic matter. Its technical core is a decoupled radial/angular momentum-space quadrature, allowing massless and massive particles to be treated within one framework. Benchmarks include 1D relativistic Riemann problems, mass scans up to $L_z=42.43\,d_i$4, Bjorken-flow attractors, and vortical QGP-like flows; the reported single-GPU performance on a V100 is about 60 MLUPS on a $L_z=42.43\,d_i$5 grid with $L_z=42.43\,d_i$6, and about 20 MLUPS for $L_z=42.43\,d_i$7 (Ambrus et al., 2022).

A third branch is the graphics-oriented continuous-scale kinetic fluid simulation, which uses the lattice Boltzmann equation with a non-orthogonal central-moment-relaxation collision model, an adaptive high-order relaxation rule

$L_z=42.43\,d_i$8

and a continuous-scale mapping between samples of arbitrary resolution. The method removes the integer-ratio and aligned-boundary restrictions of classical multi-block LBM and supports automatic static and dynamic sample construction. Reported scenarios use about 1.5M to 6.8M samples, 0.4 to 2.3 seconds per iteration on GPU, and memory footprints of about 2–7 GB (Li et al., 2018). Related kinetic PDE work for chemotaxis extends the same philosophy to velocity-jump bacterial transport on Cartesian meshes in arbitrary geometries with reflective walls and WENO-type ghost-cell reconstruction, recovering aggregation, wave reflection in disks, and bar-like propagation in U-shaped channels (Filbet et al., 2013).

4. Particle-based biochemical kinetics

In biochemical reaction-diffusion, KineticSim denotes a particle-based strategy for simulating non-elementary rate laws without explicitly resolving the fast intermediate reactions that generate them. For the trimolecular effective reaction

$L_z=42.43\,d_i$9

one paper derives a proximity-based reactive region

$\mathbf B_0=B_0\hat{\mathbf e}_z$0

which encodes quasi-steady-state Michaelis–Menten-like behavior directly in the reaction geometry. Proof-of-concept simulations reproduce a Poisson steady-state population in a well-mixed birth–death test and recover the nonlinear saturation law $\mathbf B_0=B_0\hat{\mathbf e}_z$1 as $\mathbf B_0=B_0\hat{\mathbf e}_z$2 is varied (Kearney et al., 2023). The explicit claim is not that all enzyme kinetics can be handled in this way, but that a specific class reducible to the trimolecular form can be.

A later generalization adapts this construction to non-elementary bimolecular kinetics by introducing a conceptual third implicit reactant or phantom reactant. Reactive-state arrival times are sampled from an exponential law with propensity

$\mathbf B_0=B_0\hat{\mathbf e}_z$3

and the closest-approach distance is sampled so that the reaction probability becomes a function $\mathbf B_0=B_0\hat{\mathbf e}_z$4 of the nearest-partner distance. In Michaelis–Menten form this yields

$\mathbf B_0=B_0\hat{\mathbf e}_z$5

The event-driven implementation reproduces Michaelis–Menten kinetics and supports a molecule-level simulation of the Goldbeter circadian oscillator. The same work is explicit about limitations: validity requires an inverse-Laplace-transform construction for the target rate law, local well-mixedness assumptions, and non-negativity of the derived probability. For the original $\mathbf B_0=B_0\hat{\mathbf e}_z$6 Hill form in the Goldbeter model, the derived $\mathbf B_0=B_0\hat{\mathbf e}_z$7 can become negative, so it is not a valid reaction probability (Kearney et al., 31 Jul 2025). This directly addresses the common misconception that proximity-based particle simulators are restricted to elementary mass-action chemistry.

5. Kinetic Monte Carlo, electrostatics, and asymptotic-preserving transport

For lattice or off-lattice hopping models with long-range Coulomb interactions, a major KineticSim bottleneck is repeated electrostatic energy evaluation during proposal generation. An FMM-based KMC solver addresses this by exploiting the fact that each accepted KMC event is a single-particle move. Proposed moves are evaluated using local expansions plus direct interactions with the particle’s own cell and its 26 neighbors, and accepted moves update the FMM representation incrementally. For free-space boundaries, the proposal cost is

$\mathbf B_0=B_0\hat{\mathbf e}_z$8

independent of total particle number $\mathbf B_0=B_0\hat{\mathbf e}_z$9, while the accepted-move update cost is

$\beta_i=0.5$0

so the total KMC step scales linearly with $\beta_i=0.5$1 for fixed expansion order. The implementation is exposed through the PPMD framework and the coulomb_kmc Python package (Saunders et al., 2019). The point is not merely acceleration, but retention of accurate long-range electrostatics without crude truncation.

A complementary strategy appears in the 2D implementation of Kinetic-diffusion Monte Carlo in Eiron for neutral transport in highly collisional tokamak scrape-off-layer conditions. KDMC alternates an exact kinetic jump process with a Gaussian diffusive increment, so that when $\beta_i=0.5$2 the algorithm behaves like kinetic Monte Carlo, and when $\beta_i=0.5$3 the diffusive correction dominates. In 2D the diffusive displacement is sampled from a multivariate normal distribution $\beta_i=0.5$4, with $\beta_i=0.5$5 defining the diffusive flight time. Validation on slab problems shows that in the kinetic regime the 2D histograms are visually indistinguishable from the reference kinetic solution, while in the high-collision regime the runtime advantage becomes decisive: the highest-collision cases show speedups of two to three orders of magnitude (Lappi et al., 23 Sep 2025). The paper is equally explicit about current limitations, including restriction to a single domain and incomplete support for general heterogeneous backgrounds and boundary conditions.

6. KineticSim as a real-time market execution engine

The most explicit use of the name appears in “KineticSim: A Lightweight, High-Performance Execution Engine for Real-Time Market Simulators” (Jayakody et al., 19 Jun 2026). Here KineticSim is a custom CUDA engine for discrete-time uniform-price call auctions over a price grid of size $\beta_i=0.5$6, with cumulative demand and supply

$\beta_i=0.5$7

and clearing price

$\beta_i=0.5$8

Its central design pattern is persistent, state-carrying clearing for iterative multi-agent reductions: each market is mapped to one CUDA block, the entire limit-order book is kept in block shared memory across all $\beta_i=0.5$9 time steps, agent orders are aggregated through shared-memory atomics, and clearing is resolved cooperatively with parallel scans and reductions. The resulting per-step critical-path depth is reduced from $m_i/m_e=64$0 for sequential clearing to $m_i/m_e=64$1, while global-memory traffic becomes independent of the step count (Jayakody et al., 19 Jun 2026).

The implementation detail that defines this KineticSim is its on-chip persistence. The kernel is launched as $m_i/m_e=64$2, so one market corresponds to one block and one thread corresponds to one price tick. The shared-memory working set is seven $m_i/m_e=64$3-element float arrays plus one $m_i/m_e=64$4-element int array, amounting to $m_i/m_e=64$5 bytes, and the whole simulation proceeds inside one persistent kernel rather than through per-step host launches. This design yields a peak throughput of over 54.7 billion agent-events per second. On the fixed workload $m_i/m_e=64$6, $m_i/m_e=64$7, $m_i/m_e=64$8, the reported wall times are 20.2 ms for KineticSim, 68.6 s for CPU NumPy, 560.3 ms for PyTorch GPU, 862.6 ms for JAX GPU, and 169.4 ms for a naive CUDA baseline, corresponding to speedups of 3406×, 27.8×, 42.8×, and 8.4×, respectively (Jayakody et al., 19 Jun 2026).

Correctness is evaluated in two ways. First, across 53 configurations, KineticSim and the naive custom CUDA engine produce bitwise-identical order books and clearing statistics. Second, aggregate market statistics match the CPU reference to within 0.1%. Memory use is also materially smaller than that of framework-based GPU baselines: at $m_i/m_e=64$9, the reported GPU memory footprints are 340.4 MB for PyTorch GPU, 69.06 MB for naive CUDA, and 34.63 MB for KineticSim (Jayakody et al., 19 Jun 2026). Although the concrete application is market microstructure, the paper explicitly proposes the same pattern for other iterative multi-agent workloads with compact shared state and repeated localized reductions.

7. Cross-domain themes and interpretive significance

Taken together, the literature indicates that KineticSim is not best defined by application domain, but by a recurrent computational logic. One recurrent pattern is regime bridging: KDMC moves continuously between kinetic and diffusive limits (Lappi et al., 23 Sep 2025); the relativistic lattice kinetic solver spans fluid and rarefied-gas regimes (Ambrus et al., 2022); the gas-kinetic turbulence scheme approaches Navier–Stokes behavior at low rarefaction and activates kinetic corrections when scale separation deteriorates (Righi, 2013). Another is state localization: local magnetic-field organization governs anisotropy in kinetic Alfvén turbulence (Groselj et al., 2017), rational-surface spacing localizes KIM near minimum magnetic shear (Li et al., 2024), and block-local shared memory is the decisive state container in the market engine (Jayakody et al., 19 Jun 2026). A third is implicit representation of fast processes: enzyme-binding transients are folded into moving reaction boundaries or phantom-reactant probabilities rather than simulated event by event (Kearney et al., 2023, Kearney et al., 31 Jul 2025).

This suggests two clarifications. First, KineticSim should not be conflated with any single solver family such as PIC, LBM, gyrokinetics, or KMC. Second, “kinetic” does not uniformly mean higher fidelity than fluid or agent-based alternatives; in several cases it instead denotes a more appropriate state representation for non-local transport, rarefaction, multiscale reaction kinetics, or reduction-heavy shared-state simulation. Within that broader interpretation, the explicit market-simulation engine named KineticSim is one specialized instance of a wider kinetic-simulation tradition rather than an isolated terminological anomaly (Jayakody et al., 19 Jun 2026).