arXiv subjects

Kun Xu

Publications and source records attributed to Kun Xu.

At least 271 records · Page 15Linked to original sources

arXiv2020

A three-dimensional unified gas-kinetic wave-particle solver for flow computation in all regimes

In this paper, the unified gas-kinetic wave-particle (UGKWP) method has been constructed on three-dimensional unstructured mesh with parallel computing for multiscale flow simulation. Following the direct modeling methodology of the unified gas-kinetic scheme (UGKS), the UGKWP method models the flow dynamics uniformly in different regime and gets the local cell's Knudsen number dependent numerical solution directly without the requirement of kinetic scale cell resolution. The UGKWP method is composed of evolution of deterministic wave and stochastic particles. With the dynamic wave-particle decomposition, the UGKWP method is able to capture the continuum wave interaction and rarefied particle transport under a unified framework and achieves the high efficiency in different flow regime. The UGKWP flow solver is validated by many three-dimensional test cases of different Mach and Knudsen numbers, which include 3D shock tube problem, lid-driven cavity flow, high-speed flow passing through a cubic object, and hypersonic flow around a space vehicle. The parallel performance has been tested on the Tianhe-2 supercomputer, and reasonable parallel performance has been observed up to one thousand core processing. Due to wave-particle formulation, the UGKWP method has great potential in solving three-dimensional multiscale transport with the co-existence of continuum and rarefied flow regimes, especially for the high-speed rarefied and continuum flow around space vehicle in near space flight.

physics.comp-ph↗

arXiv2020

Understanding and Stabilizing GANs' Training Dynamics with Control Theory

Generative adversarial networks (GANs) are effective in generating realistic images but the training is often unstable. There are existing efforts that model the training dynamics of GANs in the parameter space but the analysis cannot directly motivate practically effective stabilizing methods. To this end, we present a conceptually novel perspective from control theory to directly model the dynamics of GANs in the function space and provide simple yet effective methods to stabilize GANs' training. We first analyze the training dynamic of a prototypical Dirac GAN and adopt the widely-used closed-loop control (CLC) to improve its stability. We then extend CLC to stabilize the training dynamic of normal GANs, where CLC is implemented as a squared $L2$ regularizer on the output of the discriminator. Empirical results show that our method can effectively stabilize the training and obtain state-of-the-art performance on data generation tasks.

cs.LG↗

arXiv2020

Learning Implicit Generative Models by Teaching Explicit Ones

Implicit generative models are difficult to train as no explicit density functions are defined. Generative adversarial nets (GANs) present a minimax framework to train such models, which however can suffer from mode collapse due to the nature of the JS-divergence. This paper presents a learning by teaching (LBT) approach to learning implicit models, which intrinsically avoids the mode collapse problem by optimizing a KL-divergence rather than the JS-divergence in GANs. In LBT, an auxiliary density estimator is introduced to fit the implicit model's distribution while the implicit model teaches the density estimator to match the data distribution. LBT is formulated as a bilevel optimization problem, whose optimal generator matches the true data distribution. LBT can be naturally integrated with GANs to derive a hybrid LBT-GAN that enjoys complimentary benefits. Finally, we present a stochastic gradient ascent algorithm with unrolling to solve the challenging learning problems. Experimental results demonstrate the effectiveness of our method.

cs.LG↗

arXiv2020

A Unified Gas-kinetic Scheme for Micro Flow Simulation Based on Linearized Kinetic Equation

The flow regime of micro flow varies from collisionless regime to hydrodynamic regime according to the Knudsen number. On the kinetic scale, the dynamics of micro flow can be described by the linearized kinetic equation. In the continuum regime, hydrodynamic equations such as linearized Navier-Stokes equations and Euler equations can be derived from the linearized kinetic equation by the Chapman-Enskog asymptotic analysis. In this paper, based on the linearized kinetic equation we are going to propose a unified gas kinetic scheme scheme (UGKS) for micro flow simulation, which is an effective multiscale scheme in the whole micro flow regime. The important methodology of UGKS is the following. Firstly, the evolution of microscopic distribution function is coupled with the evolution of macroscopic flow quantities. Secondly, the numerical flux of UGKS is constructed based on the integral solution of kinetic equation, which provides a genuinely multiscale and multidimensional numerical flux. The UGKS recovers the linear kinetic solution in the rarefied regime, and converges to the linear hydrodynamic solution in the continuum regime. An outstanding feature of UGKS is its capability of capturing the accurate viscous solution even when the cell size is much larger than the kinetic kinetic length scale, such as the capturing of the viscous boundary layer with a cell size ten times larger than the particle mean free path. Such a multiscale property is called unified preserving (UP) which has been studied in \cite{guo2019unified}. In this paper, we are also going to give a mathematical proof for the UP property of UGKS.

physics.comp-ph↗

arXiv2020

Three dimensional high-order gas-kinetic scheme for supersonic isotropic turbulence II: coarse-grained analysis of compressible $K_{sgs}$ budget

The direct numerical simulation (DNS) of compressible isotropic turbulence up to the supersonic regime $Ma_{t} = 1.2$ has been investigated by high-order gas-kinetic scheme (HGKS) [{\it{Computers}} \& {\it{Fluids, 192, 2019}}]. In this study, the coarse-grained analysis of subgrid-scale (SGS) turbulent kinetic energy $K_{sgs}$ budget is fully analyzed for constructing one-equation SGS model in the compressible large eddy simulation (LES). The DNS on a much higher turbulent Mach number up to $Ma_{t} = 2.0$ has been obtained by HGKS, which confirms the super robustness of HGKS. Then, the exact compressible SGS turbulent kinetic energy $K_{sgs}$ transport equation is derived with density weighted filtering process. Based on the compressible $K_{sgs}$ transport equation, the coarse-grained processes are implemented on three sets of unresolved grids with the Box filter. The coarse-grained analysis of compressible $K_{sgs}$ budgets shows that all unresolved source terms are dominant terms in current system. Especially, the magnitude of SGS pressure-dilation term is in the order of SGS solenoidal dissipation term within the initial acoustic time scale. Therefore, it can be concluded that the SGS pressure-dilation term cannot be neglected as the previous work. The delicate coarse-grained analysis of SGS diffusion terms in compressible $K_{sgs}$ equation confirms that both the fluctuation velocity triple correlation term and the pressure-velocity correlation term are dominant terms. Current coarse-grained analysis gives an indication of the order of magnitude of all SGS terms in compressible $K_{sgs}$ budget, which provides a solid basis for compressible LES modeling in high Mach number turbulent flow.

physics.flu-dyn↗

arXiv2020

High-order gas-kinetic scheme with parallel computation for direct numerical simulation of turbulent flows

The performance of high-order gas-kinetic scheme (HGKS) has been investigated for the direct numerical simulation (DNS) of isotropic compressible turbulence up to the supersonic regime. Due to the multi-scale nature and coupled temporal-spatial evolution process, HGKS provides a valid tool for the numerical simulation of compressible turbulent flow. Based on the domain decomposition and message passing interface (MPI), a parallel HGKS code is developed for large-scale computation in this paper. The standard tests from the nearly incompressible flow to the supersonic one, including Taylor-Green vortex problem, turbulent channel flow and isotropic compressible turbulence, are presented to validate the parallel scalability, efficiency, accuracy and robustness of parallel implementation. The performance of HGKS for the nearly incompressible turbulence is comparable with the high-order finite difference scheme, including the resolution of flow structure and efficiency of computation. Based on the accuracy of the numerical solution, the numerical dissipation of the scheme in the turbulence simulation is quantitatively evaluated. As a mesoscopic method, HGKS performs better than both lattice Boltzmann method (LBM) and discrete unified gas-kinetic scheme (DUGKS), due to its high-order accuracy. Meanwhile, based on the kinetic formulation HGKS shows advantage for supersonic turbulent flow simulation with its accuracy and robustness. The current work demonstrates the capability of HGKS as a powerful DNS tool from the low speed to supersonic turbulence study, which is less reported under the framework of finite volume scheme.

physics.comp-ph↗

arXiv2020

Star Formation in Massive Galaxies at Redshift $z \sim 0.5$

It is believed that massive galaxies have quenched their star formation because of active galactic nucleus feedback. However, recent studies have shown that some massive galaxies are still forming stars. We analyze the morphology of star formation regions for galaxies of stellar mass larger than 10$^{11.3}$ M$_{\odot}$ at around redshift $z_r=0.5$ using $u-z$ color images. We find that about $20\%$ of the massive galaxies are star-forming (SF) galaxies, and most of them ($\sim 85\%$) have asymmetric structures induced by recent mergers. Moreover, for these asymmetric galaxies, we find that the asymmetry of the SF regions becomes larger for bluer galaxies. Using the Illustris simulation, we can qualitatively reproduce the observed relation between asymmetry parameter and color. Furthermore, using the merger trees in the simulation, we find a correlation between the color of the main branch galaxies at $z_r=0.5$ and the sum of the Star Formation Rates (SFRs) of the recently accreted galaxies, which implies that star formation of the accreted galaxies has contributed to the observed star formation of the massive (host) galaxies (ex situ star formation). Furthermore, we find two blue and symmetric galaxies, candidates for massive blue disks, in our observed sample, which indicates that about $\sim 10\%$ of massive SF galaxies are forming stars in the normal mode of disk star formation (in situ star formation). With the simulation, we find that the disk galaxies at $z_r \approx 0.5$ should have experienced few major mergers during the last 4.3 Gyrs.

astro-ph.GA↗

arXiv2020

Extracting the magnitude of magnetic field at freeze-out in heavy-ion collisions

A strong magnetic field influences significantly the masses of the charged light mesons. For example, the mass of charged pion increases with the magnetic field increasing. We propose this mechanism as a possible way to extract the magnitude of magnetic field at freeze-out in heavy ion collisions and thus help constrain its lifetime which is currently a major open question to resolve. Specifically we show that the ratio between the yield of charged pions and that of charged rhos is very sensitive to the magnetic field value at freeze-out. By using a viscous-hydrodynamic framework (iEBE-VISHNU) to simulate heavy ion collisions and implementing magnetic-field-dependent meson masses, we compute their yields and predict the dependence of such ratio on the magnetic field. We suggest to use this ratio of charged rho yield over charged pion yield as an experimental observable to extract the possible magnetic field at freeze-out in heavy ion collisions.

hep-ph↗

arXiv2020

On the Role of Conceptualization in Commonsense Knowledge Graph Construction

Commonsense knowledge graphs (CKGs) like Atomic and ASER are substantially different from conventional KGs as they consist of much larger number of nodes formed by loosely-structured text, which, though, enables them to handle highly diverse queries in natural language related to commonsense, leads to unique challenges for automatic KG construction methods. Besides identifying relations absent from the KG between nodes, such methods are also expected to explore absent nodes represented by text, in which different real-world things, or entities, may appear. To deal with the innumerable entities involved with commonsense in the real world, we introduce to CKG construction methods conceptualization, i.e., to view entities mentioned in text as instances of specific concepts or vice versa. We build synthetic triples by conceptualization, and further formulate the task as triple classification, handled by a discriminatory model with knowledge transferred from pretrained language models and fine-tuned by negative sampling. Experiments demonstrate that our methods can effectively identify plausible triples and expand the KG by triples of both new nodes and edges of high diversity and novelty.

cs.CL↗

arXiv2020

An acoustic and shock wave capturing compact high-order gas-kinetic scheme with spectral-like resolution

In this paper, a compact high-order gas-kinetic scheme (GKS) with spectral resolution will be presented and used in the simulation of acoustic and shock waves. For accurate simulation, the numerical scheme is required to have excellent dissipation-dispersion preserving property, while the wave modes, propagation characteristics, and wave speed of the numerical solution should be kept as close as possible to the exact solution of governing equations. For compressible flow simulation with shocks, the numerical scheme has to be equipped with proper numerical dissipation to make a crispy transition in the shock layer. Based on the high-order gas evolution model, the GKS provides a time accurate solution at a cell interface, from which both time accurate flux function and the time evolving flow variables can be obtained. The GKS updates explicitly both cell-averaged conservative flow variables and the cell-averaged gradients by applying Gauss-theorem along the boundary of the control volume. Based on the cell-averaged flow variables and cell-averaged gradients, a reconstruction with compact stencil can be obtained. With the same stencil of a second-order scheme, a reconstruction up to 8th-order spacial accuracy can be constructed, which include the nonlinear and linear reconstruction for the non-equilibrium and equilibrium states respectively. The GKS unifies the nonlinear and linear reconstruction through a time evolution process at a cell interface from the non-equilibrium state to an equilibrium one. In the region between these two limits, the contribution from nonlinear and linear reconstructions depends on the weighting functions of $\exp(-Δt/τ)$ and $(1-\exp(-Δt /τ))$, where $Δt$ is the time step and $τ$ is the particle collision time, which is enhanced in the shock region. As a result, both shock and acoustic wave can be captured accurately in GKS.

physics.comp-ph↗

arXiv2020

Rethinking Softmax Cross-Entropy Loss for Adversarial Robustness

Previous work shows that adversarially robust generalization requires larger sample complexity, and the same dataset, e.g., CIFAR-10, which enables good standard accuracy may not suffice to train robust models. Since collecting new training data could be costly, we focus on better utilizing the given data by inducing the regions with high sample density in the feature space, which could lead to locally sufficient samples for robust learning. We first formally show that the softmax cross-entropy (SCE) loss and its variants convey inappropriate supervisory signals, which encourage the learned feature points to spread over the space sparsely in training. This inspires us to propose the Max-Mahalanobis center (MMC) loss to explicitly induce dense feature regions in order to benefit robustness. Namely, the MMC loss encourages the model to concentrate on learning ordered and compact representations, which gather around the preset optimal centers for different classes. We empirically demonstrate that applying the MMC loss can significantly improve robustness even under strong adaptive attacks, while keeping state-of-the-art accuracy on clean inputs with little extra computation compared to the SCE loss.

cs.LG↗

arXiv2020

Mixup Inference: Better Exploiting Mixup to Defend Adversarial Attacks

It has been widely recognized that adversarial examples can be easily crafted to fool deep networks, which mainly root from the locally non-linear behavior nearby input examples. Applying mixup in training provides an effective mechanism to improve generalization performance and model robustness against adversarial perturbations, which introduces the globally linear behavior in-between training examples. However, in previous work, the mixup-trained models only passively defend adversarial attacks in inference by directly classifying the inputs, where the induced global linearity is not well exploited. Namely, since the locality of the adversarial perturbations, it would be more efficient to actively break the locality via the globality of the model predictions. Inspired by simple geometric intuition, we develop an inference principle, named mixup inference (MI), for mixup-trained models. MI mixups the input with other random clean samples, which can shrink and transfer the equivalent perturbation if the input is adversarial. Our experiments on CIFAR-10 and CIFAR-100 demonstrate that MI can further improve the adversarial robustness for the models trained by mixup and its variants.

cs.LG↗

arXiv2020

The entanglement properties of holographic QCD model with a critical end point

We investigate different entanglement properties of a holographic QCD (hQCD) model with a critical end point at finite baryon density. Firstly we consider the holographic entanglement entropy (HEE) of this hQCD model in a spherical shaped region and a strip shaped region, respectively, and find that the HEE of this hQCD model in both regions can reflect QCD phase transition. What is more is that although the area formulas and minimal area equations of the two regions are quite different, the HEE have very similar behavior on the QCD phase diagram. So we argue that the behavior of HEE on the QCD phase diagram is independent of the shape of subregions. However, as we know that HEE is not a good quantity to characterize the entanglement between different subregions of a thermal system. So we then study the mutual information (MI), conditional mutual information (CMI) and the entanglement of purification (Ep) in different strip shaped regions. We find that the three entanglement quantities have very similar behavior: their values do not change so much in the hadronic matter phase and then rise up quickly with the increase of $T$ and $\m$ in the QGP phase. Near the phase boundary, these three entanglement quantities change smoothly in the crossover region, continuously but not smoothly at CEP and show discontinuity behavior in the first phase transition region. And all of them can be used to distinguish different phases of strongly coupled matter.

hep-th↗

arXiv2020

To Relieve Your Headache of Training an MRF, Take AdVIL

We propose a black-box algorithm called {\it Adversarial Variational Inference and Learning} (AdVIL) to perform inference and learning on a general Markov random field (MRF). AdVIL employs two variational distributions to approximately infer the latent variables and estimate the partition function of an MRF, respectively. The two variational distributions provide an estimate of the negative log-likelihood of the MRF as a minimax optimization problem, which is solved by stochastic gradient descent. AdVIL is proven convergent under certain conditions. On one hand, compared with contrastive divergence, AdVIL requires a minimal assumption about the model structure and can deal with a broader family of MRFs. On the other hand, compared with existing black-box methods, AdVIL provides a tighter estimate of the log partition function and achieves much better empirical results.

cs.LG↗

arXiv2020

Coordinated Reasoning for Cross-Lingual Knowledge Graph Alignment

Existing entity alignment methods mainly vary on the choices of encoding the knowledge graph, but they typically use the same decoding method, which independently chooses the local optimal match for each source entity. This decoding method may not only cause the "many-to-one" problem but also neglect the coordinated nature of this task, that is, each alignment decision may highly correlate to the other decisions. In this paper, we introduce two coordinated reasoning methods, i.e., the Easy-to-Hard decoding strategy and joint entity alignment algorithm. Specifically, the Easy-to-Hard strategy first retrieves the model-confident alignments from the predicted results and then incorporates them as additional knowledge to resolve the remaining model-uncertain alignments. To achieve this, we further propose an enhanced alignment model that is built on the current state-of-the-art baseline. In addition, to address the many-to-one problem, we propose to jointly predict entity alignments so that the one-to-one constraint can be naturally incorporated into the alignment prediction. Experimental results show that our model achieves the state-of-the-art performance and our reasoning methods can also significantly improve existing baselines.

cs.CL↗

arXiv2020

Multiplex Word Embeddings for Selectional Preference Acquisition

Conventional word embeddings represent words with fixed vectors, which are usually trained based on co-occurrence patterns among words. In doing so, however, the power of such representations is limited, where the same word might be functionalized separately under different syntactic relations. To address this limitation, one solution is to incorporate relational dependencies of different words into their embeddings. Therefore, in this paper, we propose a multiplex word embedding model, which can be easily extended according to various relations among words. As a result, each word has a center embedding to represent its overall semantics, and several relational embeddings to represent its relational dependencies. Compared to existing models, our model can effectively distinguish words with respect to different relations without introducing unnecessary sparseness. Moreover, to accommodate various relations, we use a small dimension for relational embeddings and our model is able to keep their effectiveness. Experiments on selectional preference acquisition and word similarity demonstrate the effectiveness of the proposed model, and a further study of scalability also proves that our embeddings only need 1/20 of the original embedding size to achieve better performance.

cs.CL↗

arXiv2019

Machine learning and evolutionary algorithm studies of graphene metamaterials for optimized plasmon-induced transparency

Machine learning and optimization algorithms have been widely applied in the design and optimization for photonic devices. In this article, we briefly review recent progress of this field of research and show some data-driven applications (e.g. spectrum prediction, inverse design and performance optimization) for novel graphene metamaterials (GMs). The structure of the GMs is well-designed to achieve the wideband plasmon induced transparency effect, which is regarded as optimization object and can be theoretically demonstrated by using transfer matrix method. Some classical machine learning algorithms, including k nearest neighbour, decision tree, random forest and artificial neural networks, are utilized to equivalently substitute the numerical simulation in the forward spectrum prediction and complete the inverse design for the GMs. The calculated results demonstrate that all the algorithms are effective and the random forest has advantages in terms of accuracy and training speed. Moreover, the single-objective and multi-objective optimization algorithms are used to achieve steep transmission characteristics by synthetically taking many performance metrics into consideration. The maximum difference between the transmission peaks and dips in the optimized transmission spectrum can reach 0.97. In comparison to previous works, we provide a guidance for intelligent design of photonic devices and advanced materials based on machine learning and evolutionary algorithms.

physics.optics↗

arXiv2019

Compensated magnetic insulators for extremely fast spin-orbitronics

The fast spin dynamics provide many opportunities for the future communication and memory technologies. One of the most promising examples is the domain wall (DW) racetrack memory. To achieve fast device performances, the high-speed motion of DWs is naturally demanded that leaves antiferromagnets (AFMs) and compensated ferrimagnets (FIMs) as the promising materials. While controlling and probing the dynamics of DWs in AFMs remains challenging, the fast motion of DWs with velocities around 1500 m/s has been demonstrated in metallic FIMs. The velocity of DWs in metallic FIMs is, however, suppressed by the magnetic damping of conduction electrons, which motivates us to explore how fast DWs can move in insulating FIMs where the conduction electron is absent. In this work, through synthesizing compensated FIM insulator Gd3Fe5O12 thin films with a perpendicular magnetic anisotropy, we demonstrate that the spin-orbit torque (SOT) induced motion of DWs along the Gd3Fe5O12/Pt racetrack can approach 6000 m/s. Our results show that the exceptionally fast motion of DWs can be achieved in compensated FIM insulators owing to small damping inherent to magnetic insulators, which could potentially facilitate the emerging ultrahigh-speed spintronic logics and racetrack memories by introducing insulating compensated FIMs as a new material platform.

cond-mat.mtrl-sci↗