arXiv subjects

Ziyu Huang

Publications and source records attributed to Ziyu Huang.

At least 19 recordsLinked to original sources

arXiv2026

Weave: Fine-Grained Dynamic SM Scheduling in an MoE Megakernel for Compute-Communication Overlap

Mixture-of-Experts (MoE) inference under expert parallelism (EP) turns each MoE layer into a distributed computation with costly dispatch and combine communication. State-of-the-art systems reduce this cost through communication-computation overlap, splitting the GPU's SMs for communication and computation respectively. However, this approach still leaves GPU resources wasted along two dimensions. Spatially, the best SM split is determined by each layer's routing result and varies across layers and GPUs, so fixed policies mismatch the workload and waste either NVLink bandwidth or compute throughput. Temporally, complex MoE data dependencies introduce bubbles that leave SMs idle. We present Weave, to our knowledge the first MoE overlap system that performs fine-grained dynamic SM scheduling - deciding per layer and per GPU by routing results at runtime. Once routing completes, each layer's communication and computation volumes become known; Weave exploits this predictability through a lightweight cost model running inside the persistent megakernel: a spatial scheduler partitions SMs into communication workers and computation workers to match the communication/computation throughput ratio, and a temporal scheduler coordinates the two worker groups to minimize SM idleness. On 4x H100 SXM GPUs across six mainstream MoE models, Weave achieves a 2.89x geometric-mean MoE-layer speedup and a 1.33x geometric-mean end-to-end speedup over five state-of-the-art baselines.

cs.DC↗

arXiv2026

Accelerating Chemical Kinetics for Exoplanet Atmospheres using Neural Networks

Observations increasingly reveal the coupled radiative, chemical, and dynamical processes that shape exoplanet atmospheres. Interpreting these atmospheres requires models that can capture this complexity. However, multidimensional models remain fundamentally limited by computational cost, and answering key questions requires simulating the governing physical mechanisms at speeds classical methods cannot achieve. As a result, models often rely on simplifying approximations, such as equilibrium chemistry, even when those assumptions miss important effects. There is a pressing need for fast and accurate chemical kinetics solvers to model planetary atmospheres. Here we present a machine learning local-box chemical kinetics solver for exoplanet atmospheres using a residual flow-map architecture. We demonstrate that this surrogate model is several orders of magnitude faster than a classical solver, achieving microsecond-scale inference while retaining percent-level accuracy. The surrogate model covers a parameter space that spans $T=300$-$3000$ K, $P=10^{-6}$-$10^{4}$ bar, $Δt=10^{-3}$-$10^{8}$ s, and compositions ranging from $10^{-2}$ to $10^{3}$ times solar in both C/O ratio and metallicity. Our model outperforms several commonly used machine learning architectures and performs robustly under the extreme stiffness characteristic of atmospheric chemistry. The machine learning framework presented here is a flexible and efficient approach to emulating state-to-state flow-map problems that commonly arise in numerical simulations.

astro-ph.EP↗

arXiv2026

Simulating Ionic Liquid Fragmentation in Electrospray Thrusters with Foundation Models

Predicting the products of ionic-liquid impacts on extractor surfaces is important for electrospray-thruster lifetime analysis, yet available atomistic methods require a compromise between chemical fidelity and computational cost. Reactive force fields enable high-throughput sampling but do not explicitly resolve electronic charge redistribution and may miss relevant reaction pathways during impact, whereas mixed quantum--classical density-functional-theory molecular dynamics (DFT/MD) can capture charge redistribution and neutral-product formation at substantially higher computational cost. Pretrained atomistic foundation models have recently emerged as a potential route toward DFT-like chemical fidelity at considerably lower cost. Here, we benchmark two pretrained machine-learning interatomic potentials, MACE-MP-0 (medium) and MACE-POLAR-1, against DFT/MD and ReaxFF for geometry optimization of 1-ethyl-3-methylimidazolium tetrafluoroborate (EMI-BF$_4$) and for 10-100 eV impacts on a model Au extractor surface. The models reproduce several collision outcomes observed in DFT/MD, including ionic dissociation, high-energy covalent fragmentation, and, in particular, HF formation through neutralization-like chemistry that is not captured in the ReaxFF simulations. In the computational-performance benchmark, MACE-POLAR-1 and MACE-MP-0 (medium) completed each 2~ps trajectory in 5.12 and 2.54~min, respectively, corresponding to wall times approximately four orders of magnitude shorter than the DFT/MD reference under the reported benchmark conditions. These results support pretrained machine-learning potentials as a practical intermediate-cost approach for chemically resolved electrospray-impact simulations and motivate targeted fine-tuning with DFT data for broader applications in electrospray-thruster and electric-propulsion modeling.

physics.chem-ph↗

arXiv2026

Benchmarking Universal Machine Learning Force Fields for Molecular Dynamics of Lunar Regolith Minerals

Universal machine-learning interatomic potentials provide a promising route for accelerating molecular dynamics simulations of materials, but their transferability to lunar regolith-relevant silicates, oxides, and hydrogen-bearing surface species remains elucidated. Here, we benchmark six foundation models, MACE-MH, MatterSim, SevenNet-0, UPET, UMA, and NequIP-OAM-L, using NVT molecular dynamics simulations of four representative lunar minerals: forsterite, fayalite, ilmenite, and anorthite. Structural fidelity is evaluated using temperature stability, bond-distance statistics, bond-angle distributions, and partial radial distribution functions, with comparison to crystallographic reference data. The models reproduce Si--O, Mg--O, Al--O, and Ca--O local environments reasonably well, while Fe--O and Ti--O coordination environments show broader distributions and larger short-timescale fluctuations, highlighting the need for further validation and fine tuning with additional ground truth data for Fe- and Ti-bearing lunar phases. Hydroxylated surface tests show consistent O--H bond-distance distributions across models and minerals, suggesting that these foundation models may provide useful starting points for screening surface hydroxyl stability and volatile-related processes. Performance benchmarks on a single NVIDIA RTX 4090 show that SevenNet-0, MatterSim, and UPET provide the highest throughput among the six tested models, MACE-MH remains practical at intermediate cost, and UMA and NequIP-OAM-L extend the comparison to newer foundation potentials at higher runtime cost and memory demand. These results provide an initial benchmark for applying universal foundation models to lunar mineral simulations and identify key directions for future ab initio validation, model fine-tuning, and applications to lunar volatile evolution, space weathering, ISRU, and polar sample return studies.

physics.geo-ph↗

arXiv2026

Creation of Lunar-Like Rims in Ilmenite using Synthetic Solar Wind

Space weathering of lunar minerals, due to bombardment from solar wind (SW) particles and micrometeoroid impacts, modifies the mineralogy within tens of nanometers of the surface, i.e., the rim. Spectroscopic signatures of these modifications, observed via remote sensing, have long been used to gauge surface exposure times on the Moon. However, the relative contributions of SW and micrometeoroids in the creation of rim features are still debated, particularly for the nanometer-scale clusters known as nanophase iron (npFe0), which commonly form in ferrous minerals. We address this issue in the laboratory, using deuterium ions and low-energy electrons as a synthetic solar wind plasma to irradiate ilmenite (FeTiO3), a common lunar mineral. Characterization by high-resolution scanning transmission electron microscopy and electron energy-loss spectroscopy shows that the SW alone creates rims with all the main characteristics of lunar samples. We conclusively identify npFe0 and quantify its distribution as a function of depth and fluence, allowing us to estimate the SW exposure of Apollo soil 71501. Our results confirm that small npFe0 particles (<10 nm in diameter) form from SW irradiation. Such experiments provide microscopic details of space weathering, improving the link between surface modification processes and macroscopic remote-sensing data.

astro-ph.EP↗

arXiv2026

The Internal Nebular Attenuation Curve of Three-Dimensional Turbulent HII regions

The internal dust attenuation of the Hii region reduces the observed emission-line fluxes. Turbulent density fields within each Hii region change the degree of the line-of-the-sight obscuration of the emission-line fluxes. In this paper, we implement the dust Monte-Carlo radiative transfer in the latest M3D code, creating the emission-line maps attenuated by the internal turbulent dust obscuration with the varying Mach numbers. The internal density and temperature fluctuations of Hii regions make the radiative transfer of hydrogen lines neither Case A nor Case B conditions, resulting in the global Hα to H\b{eta} ratio of approximately 3.02-3.03, differing from the widely-used value of 2.86. This deviation from Case B is because the temperature of these Hii regions is cooler than 10,000 K. We further derive the internal nebular attenuation curve from the attenuated Hydrogen lines, finding that the clumpy structures within Hii regions do not change the slope of the internal attenuation curve. This is because the heavy dust obscuration of dense clumps is canceled out by the high in-situ production of emission-line intensities.

astro-ph.GA↗

arXiv2026

Coupled Space Weathering: Nanophase Iron Formation by Micrometeoroid Impact and Solar Wind Sputtering

Understanding the interplay between micrometeoroid impacts and solar wind irradiation is crucial for interpreting lunar surface evolution. Using reactive molecular dynamics simulations and surface binding energy (SBE) analyses, this study investigates the coupled effects of these two dominant space weathering processes on lunar regolith composed of Fe$_2$SiO$_4$. Our simulations reveal that micrometeoroid impacts significantly modify the lunar surface, creating structurally heterogeneous zones with varying SBEs across microcrater morphologies. Specifically, microcrater floors exhibit enhanced surface cohesion due to high-density compaction, whereas microcrater walls and ejecta show weakened structures. Applying Sigmund's sputtering theory with these SBEs indicates differential sputtering yields for Fe, Si, and O suggesting preferential retention of heavier elements like Fe. This selective sputtering mechanism supports the formation and growth of nanophase metallic iron (npFe$^0$) clusters, influencing the optical and compositional maturation of the lunar surface. These findings advance our understanding of lunar space weathering processes.

astro-ph.EP↗

arXiv2026

Revealing Exotic Nanophase Iron in Lunar Samples Through Impact-Driven Spatial Fingerprints

Nanophase iron (npFe) plays a crucial role in controlling the optical, chemical, and physical evolution of lunar regolith grains. While in-situ formation of npFe via reduction of native Fe-bearing minerals has long been considered a dominant pathway, recent mineralogical evidence from X.Zeng et al. (2025) reveals that the source of a significant fraction of npFe may be delivered directly by exotic micrometeoroid impacts (exotic npFe). Yet the atomic-scale processes governing how exotic np-Fe forms and survives during hypervelocity impacts remain largely unknown. To quantitatively compare in-situ and exotic delivery and formation of npFe, we perform a series of innovative atomistic modeling of micrometeoroid impacts with distinct projectile target compositions: (1) SiO$_2$ projectiles on Fe$_2$SiO$_4$ targets (in-situ formation), (2) Fe$_2$SiO$_4$ projectiles on SiO$_2$ targets (exotic delivery). Our results reveal distinct mechanistic fingerprints: in-situ np-Fe forms diffusely and radially around the impact site, whereas exotic np-Fe is efficiently retained and concentrated in asymmetric, momentum-aligned clusters. These contrasting spatial signatures provide a potential diagnostic criterion for distinguishing exotic versus in-situ np-Fe in returned lunar soils. In agreement with Chang'e-5 observations, our simulations demonstrate that exotic np-Fe production can be substantial, particularly in Fe-poor terrains such as highland regions. These findings highlight the need to account for exotic np-Fe when interpreting space weathering processes and remote-sensing data for the Moon and other airless bodies.

astro-ph.EP↗

arXiv2026

RPO: Fine-Tuning Visual Generative Models via Rich Vision-Language Preferences

Traditional preference tuning methods for LLMs/Visual Generative Models often rely solely on reward model labeling, which can be opaque, offer limited insights into the rationale behind preferences, and are prone to issues such as reward hacking or overfitting. We introduce Rich Preference Optimization (RPO), a novel pipeline that leverages rich feedback signals from Vision Language Models (VLMs) to improve the curation of preference pairs for fine-tuning visual generative models like text-to-image diffusion models. Our approach begins with prompting VLMs to generate detailed critiques of synthesized images, from which we further prompt VLMs to extract reliable and actionable image editing instructions. By implementing these instructions, we create refined images, resulting in synthetic, informative preference pairs that serve as enhanced tuning datasets. We demonstrate the effectiveness of our pipeline and the resulting datasets in fine-tuning state-of-the-art diffusion models.

cs.LG↗

arXiv2026

A Class of Degenerate Mean Field Games, Associated FBSDEs and Master Equations

In this paper, we study a class of degenerate mean field games (MFGs) with state-distribution dependent and unbounded functional diffusion coefficients. With a probabilistic method, we study the well-posedness of the forward-backward stochastic differential equations (FBSDEs) associated with the MFG and arising from the maximum principle, and estimate the corresponding Jacobian and Hessian flows. We further establish the classical regularity of the value functional $V$; in particular, we show that when the cost function is $C^3$ in the spatial and control variables and $C^2$ in the distribution argument, then the value functional is $C^1$ in time and $C^2$ in the spatial and distribution variables. As a consequence, the value functional $V$ is the unique classical solution of the degenerate MFG master equation.

math.OC↗

arXiv2026

Channel Knowledge Map Construction via Guided Flow Matching

The efficient construction of accurate channel knowledge maps (CKMs) is crucial for unleashing the full potential of environment-aware wireless networks, yet it remains a difficult ill-posed problem due to the sparsity of available location-specific channel knowledge data. Although diffusion-based methods such as denoising diffusion probabilistic models (DDPMs) have been exploited for CKM construction, they rely on iterative stochastic sampling, rendering them too slow for real-time wireless applications. To bridge the gap between high fidelity and efficient CKM construction, this letter introduces a novel framework based on linear transport guided flow matching (LT-GFM). Deviating from the noise-removal paradigm of diffusion models, our approach models the CKM generation process as a deterministic ordinary differential equation (ODE) that follows linear optimal transport paths, thereby drastically reducing the number of required inference steps. We propose a unified architecture that is applicable to not only the conventional channel gain map (CGM) construction, but also the more challenging spatial correlation map (SCM) construction. To achieve physics-informed CKM constructions, we integrate environmental semantics (e.g., building masks) for edge recovery and enforce Hermitian symmetry for property of the SCM. Simulation results verify that LT-GFM achieves superior distributional fidelity with significantly lower Fréchet Inception Distance (FID) and accelerates inference speed by a factor of 25 compared to DDPMs.

cs.IT↗

arXiv2025

FlashFuser: Expanding the Scale of Kernel Fusion for Compute-Intensive Operators via Inter-Core Connection

The scaling of computation throughput continues to outpace improvements in memory bandwidth, making many deep learning workloads memory-bound. Kernel fusion is a key technique to alleviate this problem, but the fusion strategies of existing compilers and frameworks are limited to using local scratchpad memory. When the intermediate results exceed the limited capacity (such as FFN), the fusion fails. Although modern GPUs (like the NVIDIA H100) now incorporate an inter-core connection mechanism known as Distributed Shared Memory(DSM)--providing a larger, high-bandwidth, and low-latency on-chip memory pool--this hardware potential has yet to be exploited by software frameworks. To bridge this gap, we present FlashFuser, the first compiler framework to utilize inter-core connection for kernel fusion on modern GPUs. FlashFuser extends established fusion techniques to the DSM domain through three core contributions. First, we propose a powerful DSM-based communication abstraction that formalizes complex cluster-based data exchange patterns, such as reduce, shuffle and multiply. Second, we introduce a dataflow analyzer that generalizes loop scheduling, resource mapping, and tile selection to the distributed memory hierarchy; it determines the optimal execution order and tile sizes by quantifying data movement across memory levels. Finally, FlashFuser integrates these components into a unified search engine that employs analytical cost modeling and DSM-aware pruning strategies to efficiently discover the optimal execution plan. Our evaluation on an NVIDIA H100 GPU shows that FlashFuser reduces memory access by 58% and delivers kernel speedups of 3.3x against highly-tuned libraries and 4.1x against state-of-the-art compilers, resulting in a 1.24x end-to-end speedup.

cs.DC↗

arXiv2025

A Variational Approach to Mean Field Type Control

Variational methods have been used to study stochastic control for long, see Bensoussan (1982) and Bensoussan-Lions (1978) for the early works. More precisely, variational approaches apply to the study of Bellman equation as a parabolic quasi-linear equation, when the nonlinearity affects only the gradient of the solution, and the second order derivative term is linear and not degenerate. This corresponds to a stochastic control problem, where the state equation is a diffusion process. The primary objective of this article is to extend this approach to mean field control theory, as an alternative to the current approach, which considers a coupled system of Hamilton-Jacobi (HJ) and Fokker-Planck (FP) equations, since the introduction of the theory by Lasry-Lions (2007). The main novelty lies in that the equation studied here is the HJB equation, neither the HJ-FP system nor the master equation; and our results also provide another perspective for probabilistic approaches; see Chassagneux-Crisan-Delarue (2022), Bensoussan-Wong-Yam-Yuan (2024), Bensoussan-Tai-Yam (2025) and Bensoussan-Huang-Tang-Yam (2025) for instance. Within the scope of the PDE methods, the advantage of this article is to solve a larger class of mean field control problems, with moderate regularity; and this kind of variational methods fairly require few conditions on the regularity of the coefficients.

math.OC↗

arXiv2025

Micrometeoroid Impacts: Dual Pathways for Iron Reduction and Oxidation on Lunar and Asteroidal Surfaces

Nanophase metallic iron ( $\mathrm{npFe}^0$ ) is a key indicator of space weathering on the lunar surface, primarily attributed to solar wind irradiation and micrometeoroid impacts. Recent discoveries of hematite ( $\mathrm{Fe}_2 \mathrm{O}_3$ ), a highly oxidized form of iron, in the lunar polar regions challenge the prevailing understanding of the Moon's reducing environment. This study, using ReaxFF molecular dynamics simulations of micrometeoroid impacts on fayalite ( $\mathrm{Fe}_2 \mathrm{SiO}_4$ ), investigates the atomistic mechanisms leading to both reduced and oxidized iron species. Our simulations reveals that the high-temperature and pressure conditions at the impact crater surface produces a reduced iron environment while providing a transient oxygen-rich environment in the expanding plume. Our findings bridge previously disparate observations-linking impact-driven $\mathrm{npFe}^0$ formation to the puzzling presence of oxidized iron phases on the Moon, completing the observed strong dichotomous distribution of hematite between the nearside and farside of the Moon. These findings highlight that micrometeoroid impacts, by simultaneously generating spatially distinct redox environments, provide a formation mechanism that reconciles the ubiquitous identification of nanophase metallic iron ( $\mathrm{npFe}^0$ ) in returned lunar samples with $\mathrm{Fe}^{3+}$ signatures detected by remote sensing. This underscores the dynamic nature of space weathering processes. For a more nuanced understanding of regolith evolution, we should also consider the presence of different generations or types of $\mathrm{npFe}{ }^0$, such as those formed from solar wind reduction versus impact disproportionation.

astro-ph.EP↗

arXiv2025

Mean Field Type Control Problems Driven by Jump-diffusions

In this article, we apply a probabilistic approach to study general mean field type control (MFTC) problems with jump-diffusions, and give the first global-in-time solution. We allow the drift coefficient $b$ and the diffusion coefficient $σ$ to nonlinearly depend on the state, distribution and control variables, and both can be unbounded and possibly degenerate; besides, the jump coefficient $γ$ is allowed to be non-constant. To tackle the non-linear and control-dependent diffusion $σ$, we further formulate a joint cone property and estimates for both processes $P$ and $Q$ of the corresponding adjoint process (where $(P,Q,R)$ is the solution triple of the associated adjoint process as a backward stochastic differential equation with jump), in contrast to our previous single cone property of the only process $P$. We study first the system of forward-backward stochastic differential equations (FBSDEs) with jumps arising from the maximum principle, and then the related Jacobian flows, which altogether yield the classical regularity of the value function and thus allow us to show that the value function is the unique classical solution of the HJB integro-partial differential equation. Most importantly, our proposed probabilistic approach can apparently handle the MFTC problem driven by a fairly general process far beyond Brownian motion, in a relatively easier manner than the existing analytic approach.

math.PR↗

arXiv2025

On Convergence Rates of General $N$-Player Stackelberg Games to their Mean Field Limits

In this article, we establish precise convergence rates of a general class of $N$-Player Stackelberg games to their mean field limits, which allows the response time delay of information, empirical distribution based interactions, and the control-dependent diffusion coefficients. All these features makes our problem nonstandard, barely been touched in the literature, and they complicate the analysis and therefore reduce the convergence rate. We first justify the same convergence rate for both the followers and the leader. Specifically, for the most general case, the convergence rate is shown to be $\mathcal{O}\left(N^{-\frac{2(q-2)}{n_1(3q-4)}}\right)$ when $n_1>4$ where $n_1$ is the dimension of the follower's state, and $q$ is the order of the integration of the initial; and this rate has yet been shown in the literature, to the best of our knowledge. Moreover, by classifying cases according to the state dimension $n_1$, the nature of the delay, and the assumptions of the coefficients, we provide several subcases where faster convergence rates can be obtained; for instance the $\mathcal{O}\left(N^{-\frac{2}{3n_1}}\right)$-convergence when the diffusion coefficients are independent of control variable. Our result extends the standard $o(1)$-convergence result for the mean field Stackelberg games in the literature, together with the $\mathcal{O}(N^{-\frac{1}{n_1+4}})$-convergence for the mean field games with major and minor players. We also discuss the special case where our coefficients are linear in distribution argument while nonlinear in state and control arguments, and we establish an $\mathcal{O}(1/\sqrt{N})$ convergence rate, which extends the linear quadratic cases in the literature.

math.OC↗

arXiv2025

RoboSeek: You Need to Interact with Your Objects

Optimizing and refining action execution through exploration and interaction is a promising way for robotic manipulation. However, practical approaches to interaction-driven robotic learning are still underexplored, particularly for long-horizon tasks where sequential decision-making, physical constraints, and perceptual uncertainties pose significant challenges. Motivated by embodied cognition theory, we propose RoboSeek, a framework for embodied action execution that leverages interactive experience to accomplish manipulation tasks. RoboSeek optimizes prior knowledge from high-level perception models through closed-loop training in simulation and achieves robust real-world execution via a real2sim2real transfer pipeline. Specifically, we first replicate real-world environments in simulation using 3D reconstruction to provide visually and physically consistent environments, then we train policies in simulation using reinforcement learning and the cross-entropy method leveraging visual priors. The learned policies are subsequently deployed on real robotic platforms for execution. RoboSeek is hardware-agnostic and is evaluated on multiple robotic platforms across eight long-horizon manipulation tasks involving sequential interactions, tool use, and object handling. Our approach achieves an average success rate of 79%, significantly outperforming baselines whose success rates remain below 50%, highlighting its generalization and robustness across tasks and platforms. Experimental results validate the effectiveness of our training framework in complex, dynamic real-world settings and demonstrate the stability of the proposed real2sim2real transfer mechanism, paving the way for more generalizable embodied robotic learning. Project Page: https://russderrick.github.io/Roboseek/

cs.RO↗

arXiv2025

Surrogate Modeling of Landau Damping with Deep Operator Networks

Kinetic simulations excel at capturing microscale plasma physics phenomena with high accuracy, but their computational demands make them impractical for modeling large-scale space and astrophysical systems. In this context, we build a surrogate model, using Deep Operator Networks (DeepONets), based upon the Vlasov-Poisson simulation data to model the dynamical evolution of plasmas, focusing on the Landau damping process - a fundamental kinetic phenomenon in space and astrophysical plasmas. The trained DeepONets are able to capture the evolution of electric field energy in both linear and nonlinear regimes under various conditions. Extensive validation highlights DeepONets' robust performance in reproducing complex plasma behaviors with high accuracy, paving the way for large-scale modeling of space and astrophysical plasmas.

physics.plasm-ph↗