arXiv subjects

Thomas Pfeil

Publications and source records attributed to Thomas Pfeil.

At least 19 recordsLinked to original sources

arXiv2026

Quantized Keys Steal Attention: Bias Correction for KV-Cache Compression in Video Diffusion

Chunk-wise autoregressive video diffusion models rely on a KV cache of previously generated chunks to avoid redundant computation, but this cache quickly becomes a memory bottleneck as videos grow longer. Methods that quantize the KV cache to low bitwidths reduce memory pressure but degrade video quality. We show that a key driver of this degradation is a systematic bias in attention weights: due to the convexity of the exponential in softmax attention, quantization noise inflates the contribution of cached keys, a phenomenon we call the Jensen bias. This effect causes quantized keys to steal attention mass from the unquantized current chunk. We derive a per-attention-score correction that removes this bias in expectation, computed on the fly from the quantization step sizes of the cached keys and the query norm. Using a second-order Taylor approximation, the additional computational overhead is negligible, and no additional memory is needed alongside the cache. Evaluated on MAGI-1, SkyReels-V2, and HY-WorldPlay at INT2 quantization, our correction recovers most of the quality lost to aggressive quantization, reaching near-BF16 video quality, and can outperform INT4 quantization while using 50% less memory.

cs.LG↗

arXiv2026

Turning the knobs on dust evolution: Comparing codes, parameters and their effects on planet formation and disc observables

Protoplanetary discs contain a wide range of dust sizes that influence their thermal structure and planet formation processes such as planetesimal formation and pebble accretion. Dust evolution models are therefore essential for both planet formation simulations and the interpretation of disc observations. Several open-source dust evolution codes are available, each adopting different model assumptions. We present a comparison of 1D radial simulations using (in order of complexity) two-pop-py, TriPoD, and DustPy, and 2D radial-vertical simulations with TriPoD, cuDisc, and mcdust. The comparison includes dust size distributions, dust disc masses, planetary gap structures, millimetre fluxes and disc sizes from synthetic observations, planetesimal formation regions, and planetary growth via pebble accretion. We also perform a parameter study to assess how key dust-evolution parameters influence disc evolution, planet formation, and code agreement. In 1D, two-pop-py depletes dust masses faster and produces higher dust concentrations outside planetary gaps than DustPy or TriPoD. The latter two generally agree well, except when size distributions deviate strongly from a power law. While the calculated millimetre fluxes and disc radii agree well, planetesimal formation locations and pebble accretion rates vary significantly between codes. In 2D, we compare cuDisc, mcdust, and TriPoD in simulations of turbulence- and sedimentation-driven coagulation. The dust size distributions agree well, despite the completely different numerical approaches used to model dust coagulation. The largest differences arise in the upper atmosphere, where mcdust suffers from low mass resolution and TriPoD fails to reproduce the exact shape of size distributions that deviate from a power-law.

astro-ph.EP↗

arXiv2026

Planet formation at the inner edge of the dead zone I. the interplay between accretion outbursts and dust growth

The inner edge of the dead zone in protoplanetary disks has been shown to periodically go unstable, leading to accretion outbursts and annular substructure within the dead zone. While dust opacities play a key role in this process, the thermal and dynamical effects of dust drift and growth have not been fully explored. We investigate the evolution of accretion outbursts in the inner disk and their impact on the formation of dust-rich substructure with a fully dynamic dust model. In doing so, we aim to highlight the importance and limitations of dust growth in forming planets in this region. We carry out radiation hydrodynamics simulations of a protoplanetary disk including prescriptions for the structure of the inner edge of the dead zone, viscous and irradiation heating, radiative cooling, dust-gas dynamics, and dust evolution. We find that accretion outbursts at the inner disk edge can lead to the formation of multiple dust rings that extend deep inside the dead zone (~1 au) and diffuse on viscous timescales (~10 kyr for alpha=1e-4). The rings contain dust masses of up to ~1.6 Earth masses, possibly kickstarting planet formation. Dynamic modeling of dust fragmentation enhances the total opacity during the burst, yielding more intense outbursts that penetrate deeper into the dead zone. Our results highlight the thermal and dynamical importance of treating dust dynamics self-consistently in models of accretion outbursts. Additional modeling is needed to characterize the inevitable nonaxisymmetric structures arising from accretion outbursts and their observational prospects.

astro-ph.EP↗

arXiv2025

TriPoDPy: 1D Tri-Population size distributions forDust evolution in protoplanetary disks

TriPoDPy is a code simulating the dust evolution, including dust growth and dynamics in protoplanetary disks using the parametric dust model presented in (Pfeil et al., 2024). The simulation evolves a dust distribution in a one-dimensional grid in the radial direction. It's written in Python and the core routines are implemented in Fortran90. The code not only solves for the evolution of the dust but also the gas disk with the canonical alpha-description (Shakura & Sunyaev, 1973). In addition to the original model, we added descriptions of tracers for the dust and gas, which could be used for compositional tracking of additional components.

astro-ph.IM↗

arXiv2025

Fragmentation-limited dust filtration in 2D simulations of planet-disk systems with dust coagulation. Parameter study and implications for the inner disk's dust mass budget and composition

Super-thermal gas giant planets or their progenitor cores are known to open deep gaps in protoplanetary disks, which stop large, drifting dust particles on their way to the inner disk. The possible separation of the disk into distinct reservoirs and the resulting dust depletion interior to the gap have important implications for planetesimal formation and the chemical and isotopic composition of the inner regions of protoplanetary disks. Dust fragmentation, however, maintains a reservoir of small grains which can traverse the gap. Dust evolution models are thus instrumental for studies of a gap's filtration efficiency. We present 2D multifluid hydrodynamic simulations of planet-disk systems with dust coagulation and fragmentation. For the first time, we evolve a series of 2D simulation with dust coagulation over 45000 planetary orbits and track the dust's size evolution and origin by using the TriPoD dust coagulation method. We investigate the effects of different planetary masses, fragmentation velocities, and viscosities on the inner disk's dust mass budget and composition, and highlight the advantages of multi-dimensional simulations over 1D models. Filtering can only be efficient for high planetary masses, high fragmentation velocities, and low diffusivities. Clear compositional distinctions between the inner and outer disk could not have been maintained by Jupiter's core if the fragmentation velocity was low, even if $\alpha \lesssim 5 \times 10^{-4}$. Significant "contamination" of the inner disk by outer-disk dust occurs in much less than $2 \times 10^5$ yr in this case and even for more massive objects. This either places tight constraints on the physical conditions in the Solar nebula or mandates consideration of alternative explanations for the NC-CC dichotomy. Astrophysical constraints on the parameters could discriminate between these possibilities.

astro-ph.EP↗

arXiv2024

Interactions Across Blocks in Post-Training Quantization of Large Language Models

Post-training quantization is widely employed to reduce the computational demands of neural networks. Typically, individual substructures, such as layers or blocks of layers, are quantized with the objective of minimizing quantization errors in their pre-activations by fine-tuning the corresponding weights. Deriving this local objective from the global objective of minimizing task loss involves two key simplifications: assuming substructures are mutually independent and ignoring the knowledge of subsequent substructures as well as the task loss. In this work, we assess the effects of these simplifications on weight-only quantization of large language models. We introduce two multi-block fine-tuning strategies and compare them against the baseline of fine-tuning single transformer blocks. The first captures correlations of weights across blocks by jointly optimizing multiple quantized blocks. The second incorporates knowledge of subsequent blocks by minimizing the error in downstream pre-activations rather than focusing solely on the quantized block. Our findings indicate that the effectiveness of these methods depends on the specific network model, with no impact on some models but demonstrating significant benefits for others.

cs.LG↗

arXiv2024

TriPoD: Tri-Population size distributions for Dust evolution. Coagulation in vertically integrated hydrodynamic simulations of protoplanetary disks

Context. Dust coagulation and fragmentation impact the structure and evolution of protoplanetary disks and set the initial conditions for planet formation. Dust grains dominate the opacities, they determine the cooling times of the gas, they influence the ionization state of the gas, and the grain surface area is an important parameter for the chemistry in protoplanetary disks. Therefore, dust evolution should not be ignored in numerical studies of protoplanetary disks. Available dust coagulation models are, however, too computationally expensive to be implemented in large-scale hydrodynamic simulations. This limits detailed numerical studies of protoplanetary disks, including these effects, mostly to one-dimensional models. Aims. We aim to develop a simple - yet accurate - dust coagulation model that can be implemented in hydrodynamic simulations of protoplanetary disks. Our model shall not significantly increase the computational cost of simulations and provide information about the local grain size distribution. Methods. The local dust size distributions are assumed to be truncated power laws. Such distributions can be characterized by two dust fluids (large and small grains) and a maximum particle size, truncating the power law. We compare our model to state-of-the-art dust coagulation simulations and calibrate it to achieve a good fit with these sophisticated numerical methods. Results. Running various parameter studies, we achieved a good fit between our simplified three-parameter model and DustPy, a state-of-the-art dust coagulation software. Conclusions. We present TriPoD, a sub-grid dust coagulation model for the PLUTO code. With TriPoD, we can perform two-dimensional, vertically integrated dust coagulation simulations on top of a hydrodynamic simulation. Studying the dust distributions in two-dimensional vortices and planet-disk systems is thus made possible.

astro-ph.EP↗

arXiv2024

Vertical shear instability with dust evolution and consistent cooling times. On the importance of the initial dust distribution

Context. Gas in protoplanetary disks mostly cools via thermal accommodation with dust particles. Thermal relaxation is thus highly sensitive to the local dust size distributions and the spatial distribution of the grains. So far, the interplay between thermal relaxation and gas turbulence has not been dynamically modeled in hydrodynamic simulations of protoplanetary disks with dust. Aims. We aim to study the effects of the vertical shear instability (VSI) on the thermal relaxation times, and vice versa. We are particularly interested in the influence of the initial dust grain size on the VSI and whether the emerging turbulence is sustained over long timescales. Results. We find that the emergence of the VSI is strongly dependent on the initial dust grain size. Coagulation also counteracts the emergence of hydrodynamic turbulence in our simulations, as shown by others before. Starting a simulation with larger grains (100 $\mu$m) generally leads to a less turbulent outcome. While the inner disk regions (within $\sim$ 70 au) develop turbulence in all three simulations, we find that the simulations with larger particles do not develop VSI in the outer disk. Conclusions. Our simulations with dynamically calculated thermal accommodation times based on the drifting and settling dust distribution show that the VSI, once developed in a disk, can be sustained over long timescales, even if grain growth is occurring. The VSI corrugates the dust layer and even diffuses the smaller grains into the upper atmosphere, where they can cool the gas. Whether the instability can emerge for a specific stratification depends on the initial dust grain sizes and the initial dust scale height. If the grains are initially $\gtrsim$ 100 $\mu$m and if the level of turbulence is initially assumed to be low, we find no VSI turbulence in the outer disk regions.

astro-ph.EP↗

arXiv2023

Dust Coagulation Reconciles Protoplanetary Disk Observations with the Vertical Shear Instability. I. Dust Coagulation and the VSI Dead Zone

Protoplanetary disks exhibit a vertical gradient in angular momentum, rendering them susceptible to the Vertical Shear Instability (VSI). The most important condition for the onset of this mechanism is a short timescale of thermal relaxation ($\lesssim 0.1$ orbital timescales). Simulations of fully VSI active disks are characterized by turbulent, vertically extended dust layers. This is in contradiction with recent observations of the outer regions of some protoplanetary disks, which appear highly settled. In this work, we demonstrate that the process of dust coagulation can diminish the cooling rate of the gas in the outer disk and extinct the VSI activity. Our findings indicate that the turbulence strength is especially susceptible to variations in the fragmentation velocity of the grains. A small fragmentation velocity of $\approx 100 \mathrm{\, cm \,s^{-1}}$ results in a fully turbulent simulation, whereas a value of $\approx 400 \mathrm{\, cm \,s^{-1}}$ results in a laminar outer disk, being consistent with observations. We show that VSI turbulence remains relatively unaffected by variations in the maximum particle size in the inner disk regions. However, we find that dust coagulation can significantly suppress the occurrence of VSI turbulence at larger distances from the central star.

astro-ph.EP↗

arXiv2022

A Neural Network Subgrid Model of the Early Stages of Planet Formation

Planet formation is a multi-scale process in which the coagulation of $\mathrm{\mu m}$-sized dust grains in protoplanetary disks is strongly influenced by the hydrodynamic processes on scales of astronomical units ($\approx 1.5\times 10^8 \,\mathrm{km}$). Studies are therefore dependent on subgrid models to emulate the micro physics of dust coagulation on top of a large scale hydrodynamic simulation. Numerical simulations which include the relevant physical effects are complex and computationally expensive. Here, we present a fast and accurate learned effective model for dust coagulation, trained on data from high resolution numerical coagulation simulations. Our model captures details of the dust coagulation process that were so far not tractable with other dust coagulation prescriptions with similar computational efficiency.

astro-ph.EP↗

arXiv2021

Hybrid SNN-ANN: Energy-Efficient Classification and Object Detection for Event-Based Vision

Event-based vision sensors encode local pixel-wise brightness changes in streams of events rather than image frames and yield sparse, energy-efficient encodings of scenes, in addition to low latency, high dynamic range, and lack of motion blur. Recent progress in object recognition from event-based sensors has come from conversions of deep neural networks, trained with backpropagation. However, using these approaches for event streams requires a transformation to a synchronous paradigm, which not only loses computational efficiency, but also misses opportunities to extract spatio-temporal features. In this article we propose a hybrid architecture for end-to-end training of deep neural networks for event-based pattern recognition and object detection, combining a spiking neural network (SNN) backbone for efficient event-based feature extraction, and a subsequent analog neural network (ANN) head to solve synchronous classification and detection tasks. This is achieved by combining standard backpropagation with surrogate gradient training to propagate gradients through the SNN. Hybrid SNN-ANNs can be trained without conversion, and result in highly accurate networks that are substantially more computationally efficient than their ANN counterparts. We demonstrate results on event-based classification and object detection datasets, in which only the architecture of the ANN heads need to be adapted to the tasks, and no conversion of the event-based input is necessary. Since ANNs and SNNs require different hardware paradigms to maximize their efficiency, we envision that SNN backbone and ANN head can be executed on different processing units, and thus analyze the necessary bandwidth to communicate between the two parts. Hybrid networks are promising architectures to further advance machine learning approaches for event-based vision, without having to compromise on efficiency.

cs.CV↗

arXiv2021

SOSP: Efficiently Capturing Global Correlations by Second-Order Structured Pruning

Pruning neural networks reduces inference time and memory costs. On standard hardware, these benefits will be especially prominent if coarse-grained structures, like feature maps, are pruned. We devise two novel saliency-based methods for second-order structured pruning (SOSP) which include correlations among all structures and layers. Our main method SOSP-H employs an innovative second-order approximation, which enables saliency evaluations by fast Hessian-vector products. SOSP-H thereby scales like a first-order method despite taking into account the full Hessian. We validate SOSP-H by comparing it to our second method SOSP-I that uses a well-established Hessian approximation, and to numerous state-of-the-art methods. While SOSP-H performs on par or better in terms of accuracy, it has clear advantages in terms of scalability and efficiency. This allowed us to scale SOSP-H to large-scale vision tasks, even though it captures correlations across all layers of the network. To underscore the global nature of our pruning methods, we evaluate their performance not only by removing structures from a pretrained network, but also by detecting architectural bottlenecks. We show that our algorithms allow to systematically reveal architectural bottlenecks, which we then remove to further increase the accuracy of the networks.

cs.LG↗

arXiv2021

High Resolution Parameter Study of the Vertical Shear Instability II: Dependence on temperature gradient and cooling time

A certain appeal to the alpha model for turbulence and related viscosity in accretion disks was that one scales the Reynolds stresses simply on the thermal pressure, assuming that turbulence driven by a certain mechanism will attain a characteristic Mach number in its velocity fluctuations. Besides the notion that there are different mechanism driving turbulence and angular momentum transport in a disk, we also find that within a single instability mechanism, here the Vertical Shear Instability, stresses do not linearly scale with thermal pressure. Here we demonstrate in numerical simulations the effect of the gas temperature gradient and the thermal relaxation time on the average stresses generated in the non-linear stage of the instability. We find that the stresses scale with the square of the exponent of the radial temperature profile at least for a range of $d \log T /d \log R = [-0.5, -1]$, beyond which the pressure scale height varies too much over the simulation domain, to provide clear results. Stresses are also dependent on thermal relaxation times, provided they are longer than $10^{-3}$ orbital periods. The strong dependence of viscous transport of angular momentum on the local conditions in the disk (especially temperature, temperature gradient, and surface density/optical depth) challenges the ideas of viscosity leading to smooth density distributions, opening a route for structure (ring) formation and time variable mass accretion.

astro-ph.EP↗

arXiv2021

ItNet: iterative neural networks with small graphs for accurate, efficient and anytime semantic segmentation

Deep neural networks have usually to be compressed and accelerated for their usage in low-power, e.g. mobile, devices. Recently, massively-parallel hardware accelerators were developed that offer high throughput and low latency at low power by utilizing in-memory computation. However, to exploit these benefits the computational graph of a neural network has to fit into the in-computation memory of these hardware systems that is usually rather limited in size. In this study, we introduce a class of network models that have a small memory footprint in terms of their computational graphs. To this end, the graph is designed to contain loops by iteratively executing a single network building block. Furthermore, the trade-off between accuracy and latency of these so-called iterative neural networks is improved by adding multiple intermediate outputs during both training and inference. We show state-of-the-art results for semantic segmentation on the CamVid and Cityscapes datasets that are especially demanding in terms of computational resources. In ablation studies, the improvement of network training by intermediate network outputs as well as the trade-off between weight sharing over iterations and the network size are investigated.

cs.LG↗

arXiv2020

The Sandwich Mode for Vertical Shear Instability in Protoplanetary Disks

Turbulence has a profound impact on the evolution of gas and dust in protoplanetary disks (PPDs), from driving the collisions and the diffusion of dust grains, to the concentration of pebbles in giant vortices, thus, facilitating planetesimal formation. The Vertical Shear Instability (VSI) is a hydrodynamic mechanism, operating in PPDs if the local rate of thermal relaxation is high enough. Previous studies of the VSI have, however, relied on the assumption of constant cooling rates, or neglected the finite coupling time between the gas particles and the dust grains. Here, we present the results of hydrodynamic simulations of PPDs with the PLUTO code that include a more realistic thermal relaxation prescription, which enables us to study the VSI in the optically thick and optically thin parts of the disk under consideration of the thermal dust-gas coupling. We show the VSI to cause turbulence even in the optically thick inner regions of PPDs in our two- and three-dimensional simulations. The collisional decoupling of dust and gas particles in the upper atmosphere and the correspondingly inefficient thermal relaxation rates lead to the damping of the VSI turbulence. Long-lived anticyclonic vortices form in our three-dimensional simulation. These structures emerge from the turbulence in the VSI-active layer, persist over hundreds of orbits and extend vertically over the whole extent of the turbulent region. We conclude that the VSI leads to turbulence and the formation of long-lived dust traps within $\pm$3 pressure scale heights distance from the disk midplane

astro-ph.EP↗

arXiv2018

Mapping the Conditions for Hydrodynamic Instability on Steady State Accretion Models of Protoplanetary Disks

Hydrodynamical instabilities in disks around young stars depend on the thermodynamic stratification of the disk and on the local rate of thermal relaxation. Here, we map the spatial extent of unstable regions for the Vertical Shear Instability (VSI), the Convective OverStability (COS), and the amplification of vortices via the Subcritical Baroclinic Instability (SBI). We use steady state accretion disk models, including stellar irradiation, accretion heating and radiative transfer. We determine the local radial and vertical stratification and thermal relaxation rate in the disk, in dependence of the stellar mass, disk mass and mass accretion rate. We find that passive regions of disks - i.e. the midplane temperature dominated by irradiation - are COS unstable about one pressure scale height above the midplane and VSI unstable at radii $> 10 \, \text{au}$. Vortex amplification via SBI should operate in most parts of active and passive disks. For active parts of disks (midplane temperature determined by accretion power) COS can become active down to the midplane. Same is true for the VSI because of the vertically adiabatic stratification of an internally heated disk. If hydro instabilities or other non-ideal MHD processes are able to create $\alpha$-stresses ($> 10^{-5}$) and released accretion energy leads to internal heating of the disk, hydrodynamical instabilities are likely to operate in significant parts of the planet forming zones in disks around young stars, driving gas accretion and flow structure formation. Thus hydro-instabilities are viable candidates to explain the rings and vortices observed with ALMA and VLT.

astro-ph.EP↗

arXiv2018

The streaming rollout of deep networks - towards fully model-parallel execution

Deep neural networks, and in particular recurrent networks, are promising candidates to control autonomous agents that interact in real-time with the physical world. However, this requires a seamless integration of temporal features into the network's architecture. For the training of and inference with recurrent neural networks, they are usually rolled out over time, and different rollouts exist. Conventionally during inference, the layers of a network are computed in a sequential manner resulting in sparse temporal integration of information and long response times. In this study, we present a theoretical framework to describe rollouts, the level of model-parallelization they induce, and demonstrate differences in solving specific tasks. We prove that certain rollouts, also for networks with only skip and no recurrent connections, enable earlier and more frequent responses, and show empirically that these early responses have better performance. The streaming rollout maximizes these properties and enables a fully parallel execution of the network reducing runtime on massively parallel devices. Finally, we provide an open-source toolbox to design, train, evaluate, and interact with streaming rollouts.

stat.ML↗

arXiv2018

Instabilities and Flow Structures in Protoplanetary Disks: Setting the Stage for Planetesimal Formation

This chapter highlights the properties of turbulence and meso-scale flow structures in protoplanetary disks and their role in the planet formation process. Here we focus on the formation of planetesimals from a gravitational collapse of a pebble cloud. Large scale and long lived flow structures - vortices and zonal flows - are a consequence of weak magneto and hydrodynamic instabilities in the pressure and entropy stratified quasi-Keplerian shear flow interacting with the fast rotation of the disk. The vortices and zonal flows on the other hand are particle traps tapping into the radial pebble flux of the disk, leading to locally sufficient accumulations to trigger gravitational collapse, directly converting pebbles to many kilometer sized planetesimals. This collapse is moderated by the streaming instability, which is a back-reaction from the particle accumulations onto the gas flow. Without trapping pebbles and increasing thus the local solid to gas ratio, this back-reaction would ultimately prevent the formation of planetsimals via turbulent diffusion. The formation of long lived flow structures is therefore a necessary condition for an efficient and fast formation of planetesimals.

astro-ph.EP↗