arXiv subjects

Vassili Kitsios

Publications and source records attributed to Vassili Kitsios.

8 recordsLinked to original sources

arXiv2026

An intercomparison of generative machine learning methods for downscaling precipitation at fine spatial scales

Machine learning (ML) offers a computationally efficient approach for generating large ensembles of high-resolution climate projections, but deterministic ML methods often smooth fine-scale structures and underestimate extremes. While stochastic generative models show promise, few studies have compared their skill under both present-day and future climates. This study compares Generative Adversarial Networks (GANs), flow matching and diffusion models across multiple configurations for downscaling daily precipitation from a regional climate model (RCM) over New Zealand. Model skill is assessed across spatial structure, distributional metrics, climatological means, extremes, ensemble calibration, and climate change signals. Unlike GANs, diffusion and flow matching models generate predictions through many sequential steps. Here we show that using higher-order differential equation solvers, the number of steps required can be reduced with only a minor reduction in skill, heavily reducing the computational burden for downscaling large ensembles, which may have otherwise prevented their use in operational settings. Overall, GANs, flow matching and diffusion perform competitively across most metrics, except that diffusion and flow matching produce higher-fidelity predictions and better-calibrated ensembles compared to GANs -which are under-dispersive. Most approaches capture mean precipitation signals reasonably well, but underestimate end-of-century climate change signals of extreme precipitation, despite being trained on RCM simulations spanning the future period. Only one GAN and one flow matching configuration can reproduce this change signal reliably. These results highlight the importance of evaluating model performance across a comprehensive set of metrics, and that neither visual realism nor good skill on standard metrics guarantee skill in predicting climate change signals.

physics.ao-ph↗

arXiv2026

Permutation-based Inference for Variational Learning of Directed Acyclic Graphs

Estimating the structure of Bayesian networks as directed acyclic graphs (DAGs) from observational data is a fundamental challenge, particularly in causal discovery. Bayesian approaches excel by quantifying uncertainty and addressing identifiability, but key obstacles remain: (i) representing distributions over DAGs and (ii) estimating a posterior in the underlying combinatorial space. We introduce PIVID, a method that jointly infers a distribution over permutations and DAGs using variational inference and continuous relaxations of discrete distributions. Through experiments on synthetic and real-world datasets, we show that PIVID can outperform deterministic and Bayesian approaches, achieving superior accuracy-uncertainty trade-offs while scaling efficiently with the number of variables.

cs.LG↗

arXiv2025

A Bayesian Ensemble Projection of Climate Change and Technological Impacts on Future Crop Yields

This paper introduces a Bayesian hierarchical modeling framework within a fully probabilistic setting for crop yield estimation, model selection, and uncertainty forecasting under multiple future greenhouse gas emission scenarios. By informing on regional agricultural impacts, this approach addresses broader risks to global food security. Extending an established multivariate econometric crop-yield model to incorporate country-specific error variances, the framework systematically relaxes restrictive homogeneity assumptions and enables transparent decomposition of predictive uncertainty into contributions from climate models, emission scenarios, and crop model parameters. In both in-sample and out-of-sample analyses focused on global wheat production, the results demonstrate significant improvements in calibration and probabilistic accuracy of yield projections. These advances provide policymakers and stakeholders with detailed, risk-sensitive information to support the development of more resilient and adaptive agricultural and climate strategies in response to escalating climate-related risks.

stat.AP↗

arXiv2024

Statistical Dynamics and Subgrid Modelling of Turbulence: From Isotropic to Inhomogeneous

Turbulence is the most important, ubiquitous, and difficult problem of classical physics. Feynman viewed it as essentially unsolved, without a rigorous mathematical basis to describe the statistical dynamics of this most complex of fluid motion. However, the paradigm shift came in 1959, with the formulation of the Eulerian direct interaction approximation (DIA) closure by Kraichnan. It was again based on renormalized perturbation theory, like quantum electrodynamics, and is a bare vertex theory that is manifestly realizable. Here, we review some of the subsequent exciting achievements in closure theory. We also document in some detail the progress that has been made in extending statistical dynamical turbulence theory to the real world of interactions with mean flows, waves and inhomogeneities such as topography. This includes numerically efficient inhomogeneous closures, like the realizable quasi-diagonal direct interaction approximation (QDIA), and even more efficient Markovian inhomogeneous closures (MICs). Recent developments include the formulation and testing of an eddy damped Markovian anisotropic closure (EDMAC) that is realizable in interactions with transient waves but is as efficient as the eddy damped quasi-normal Markovian (EDQNM). As well a similarly efficient closure, the realizable eddy damped Markovian inhomogeneous closure (EDMIC) has been developed. Moreover, we present subgrid models that cater for the complex interactions that occur in geophysical flows. Recent progress includes the determination of complete sets of subgrid terms for skilful large eddy simulations of baroclinic inhomogeneous turbulent atmospheric and oceanic flows interacting with Rossby waves and topography. The success of these inhomogeneous closures has also led to further applications in data assimilation and ensemble prediction and generalization to quantum fields.

physics.flu-dyn↗

arXiv2024

Bayesian Vector AutoRegression with Factorised Granger-Causal Graphs

We study the problem of automatically discovering Granger causal relations from observational multivariate time-series data.Vector autoregressive (VAR) models have been time-tested for this problem, including Bayesian variants and more recent developments using deep neural networks. Most existing VAR methods for Granger causality use sparsity-inducing penalties/priors or post-hoc thresholds to interpret their coefficients as Granger causal graphs. Instead, we propose a new Bayesian VAR model with a hierarchical factorised prior distribution over binary Granger causal graphs, separately from the VAR coefficients. We develop an efficient algorithm to infer the posterior over binary Granger causal graphs. Comprehensive experiments on synthetic, semi-synthetic, and climate data show that our method is more uncertainty aware, has less hyperparameters, and achieves better performance than competing approaches, especially in low-data regimes where there are less observations.

cs.LG↗

arXiv2019

Outer scaling of self-similar adverse-pressure-gradient turbulent boundary layers

The prediction of turbulent boundary layer (TBL) flow over a convex surface as in aircraft wings or gas turbine blades is a challenging problem. Finding a universal scaling law of turbulence statistics of TBLs over a wide range of adverse pressure gradients (APG) remains unresolved. Here, we introduce characteristic length and velocity scales for APG-TBLs and nondimensionalise the turbulence statistics of the recent canonical self-similar APG-TBLs by Kitsios {\it et al.} ({\it J. Fluid Mech.}, vol.829, 2018, pp. 392--419). The characteristic length scale, which is termed the `shear thickness', $δ^\ast$, is defined as the location which corresponds to the end of an actively sheared region in a turbulent shear flow, where the nondimensional shear rate normalised by the kinetic energy and the dissipation rate is approximately constant. Next, we show a universal scaling using a mixed velocity, termed the `friction-pressure velocity', $u^\ast$, which is based on total shear stress. It is revealed that the velocity fluctuations and the Reynolds stresses in TBLs over a wide range of APGs agree well with those in TBLs with zero-pressure-gradient (ZPG). The present scaling is used to scale the kinetic energy balance in TBLs, and compare them to other shear flows. Furthermore, a scaling for small-scale properties, i.e. vorticities, using $δ^\ast$ and $u^\ast$ is also obtained assuming the local equilibrium in the inertial range. The present scaling for wall-bounded shear flows, including TBLs over a wide range of pressure gradients, implies that the underlying instantaneous turbulence structures have common features under a proper scaling and is key to the development and application of turbulent models.

physics.flu-dyn↗

arXiv2017

Direct numerical simulation of a self-similar adverse pressure gradient turbulent boundary layer at the verge of separation

The statistical properties are presented for the direct numerical simulation (DNS) of a self-similar adverse pressure gradient (APG) turbulent boundary layer (TBL) at the verge of separation. The APG TBL has a momentum thickness based Reynolds number range from $Re_{δ_2}=570$ to $13800$, with a self-similar region from $Re_{δ_2} = 10000$ to $12300$. Within this domain the average non-dimensional pressure gradient parameter $β=39$, where for a unit density $β=δ_1 P_e^\prime / τ_w$, with $δ_1$ the displacement thickness, $τ_w$ the mean shear stress at the wall, and $P_e^\prime$ the farfield pressure gradient. This flow is compared to previous zero pressure gradient (ZPG) and mild APG TBL ($β=1$) results of similar Reynolds number. All flows are generated via the DNS of a TBL on a flat surface with farfield boundary conditions tailored to apply the desired pressure gradient. The conditions for self-similarity, and the appropriate length and velocity scales are derived. The mean and Reynolds stress profiles are shown to collapse when non-dimensionalised on the basis of these length and velocity scales. As the pressure gradient increases the flow has properties less like a ZPG TBL and more akin to a free shear layer.

physics.flu-dyn↗

arXiv2017

Stochastic Parameterization: Towards a new view of Weather and Climate Models

The last decade has seen the success of stochastic parameterizations in short-term, medium-range and seasonal forecasts: operational weather centers now routinely use stochastic parameterization schemes to better represent model inadequacy and improve the quantification of forecast uncertainty. Developed initially for numerical weather prediction, the inclusion of stochastic parameterizations not only provides better estimates of uncertainty, but it is also extremely promising for reducing longstanding climate biases and relevant for determining the climate response to external forcing. This article highlights recent developments from different research groups which show that the stochastic representation of unresolved processes in the atmosphere, oceans, land surface and cryosphere of comprehensive weather and climate models (a) gives rise to more reliable probabilistic forecasts of weather and climate and (b) reduces systematic model bias. We make a case that the use of mathematically stringent methods for the derivation of stochastic dynamic equations will lead to substantial improvements in our ability to accurately simulate weather and climate at all scales. Recent work in mathematics, statistical mechanics and turbulence is reviewed, its relevance for the climate problem demonstrated, and future research directions outlined.

physics.ao-ph↗