SearcharxivSearch

arXiv subjects

Philipp Huber

Publications and source records attributed to Philipp Huber.

6 recordsLinked to original sources

A preliminary exploration of the effects of baseline length for the LIFE space mission

By aiming to find and characterise dozens of habitable exoplanets through the technique of nulling interferometry, the LIFE space mission will produce transformational science. One of the key parameters for such an interferometric mission is the nulling baseline length - the distance between nulled apertures, which past studies have assumed to be 10-100m. Advances in planet occurrence statistics and simulation tools allow us now to revisit this key assumption with significantly more detail, particularly with the intention to reduce the range of baselines considered due to mission implementation concerns. We utilise the LIFEsim mission simulator along with revised mathematical tools to identify whether the range of baselines could be reduced without significantly affecting planet yield and fringe tracking performance. Along the way, we also determine a new astrophysically motivated technique for choosing which baselines are optimal for a given science target. We find that indeed, LIFE could utilise a considerably shorter range of baselines, such as 25-80m, or even discrete baselines without much (<10%) loss of performance. Nevertheless, careful trade-offs between performance and implementation simplification must be made, especially considering any spectral weighting that may be required by the scientific goals, and the potential loss of target-specific baseline optimisation.

astro-ph.IM

Energy Consumption in Parallel Neural Network Training

The increasing demand for computational resources of training neural networks leads to a concerning growth in energy consumption. While parallelization has enabled upscaling model and dataset sizes and accelerated training, its impact on energy consumption is often overlooked. To close this research gap, we conducted scaling experiments for data-parallel training of two models, ResNet50 and FourCastNet, and evaluated the impact of parallelization parameters, i.e., GPU count, global batch size, and local batch size, on predictive performance, training time, and energy consumption. We show that energy consumption scales approximately linearly with the consumed resources, i.e., GPU hours; however, the respective scaling factor differs substantially between distinct model trainings and hardware, and is systematically influenced by the number of samples and gradient updates per GPU hour. Our results shed light on the complex interplay of scaling up neural network training and can inform future developments towards more sustainable AI research.

cs.LG

Renormalization factors of quark bilinears using the DCI operator with dynamical quarks

Non-perturbative renormalization factors of bilinear quark operators are computed for the Chirally Improved lattice action with two dynamic quarks. The analysis is based on five different parameter sets with lattice size 12^3 x 24 and four parameter sets with lattice size 16^3 x 32. For the pseudoscalar renormalization factor the pion pole contribution is subtracted and chiral extrapolations are performed. Results are given in RIprime- and MSbar-scheme as well as in RGI-form.

hep-lat

Renormalization of bilinear quark operators for the chirally improved lattice Dirac operator

We compute non-perturbative renormalization constants of fermionic bilinears for the chirally improved lattice fermions in the quenched approximation of QCD. We address finite size effects and the influence of Gribov copies. Our results are presented in the RI' and MSbar schemes as well as in RGI form and we discuss relations between the renormalization constants implied by chiral symmetry. After publication we corrected the numerator of the first coefficient of α_s^3 in (24) from 3696847 to 3890527, which yields a 0.2% higher value of the conversion coefficient at μ=2 GeV.

hep-lat

Lattice calculation of low energy constants with Ginsparg-Wilson type fermions

We present a quenched lattice calculation of low energy constants using the chirally improved Dirac operator. Several lattice sizes at different lattice spacings are studied. We systematically compare various methods for computing these quantities, using pseudoscalar and axial vector correlators. We find consistent results for the different approaches, giving rise to f_π= 96(2)(4) MeV, f_K = 106(1)(8) MeV, f_K/f_π=1.11(1)(2), Sigma= -(286(4)(31) MeV)^3, the average light quark mass m = 4.1(2.4) MeV and m_s = 101(8) MeV.

hep-lat

Low energy constants from the chirally improved Dirac operator D_CI

The leading order low energy parameters like the pion decay constant or the quark condensate are well-known from "classical" low energy theorems and experiments. It is a challenge, however, to find these parameters based exclusively on an ab-initio QCD calculation. We discuss results of a quenched lattice calculation of low energy constants using the chirally improved Dirac operator. Several lattice sizes at different lattice spacings are studied, using pseudoscalar and axial vector correlators. We find consistent results for f_π= 96(2) MeV, f_K = 105(2) MeV, Σ= -(286(4) MeV)^3, the average light quark mass m = 4.1(2.4) MeV and m_s = 101(8) MeV.

hep-lat