arXiv subjects

Stefano Castruccio

Publications and source records attributed to Stefano Castruccio.

At least 19 recordsLinked to original sources

arXiv2026

Modeling Falling Backgrounds with Exponential Mixtures

Searches for new physics at the LHC often look for localized excesses on smoothly falling background distributions. Several classes of background models have been considered, including polynomials and other parametric families; however, these approaches can require extensive analysis-specific development as datasets grow. In this work, we motivate the finite exponential mixture as a flexible semi-parametric class of functions for approximating falling distributions, drawing on results from extreme value theory. Using two published datasets ($n=28,619,185$ and $n=5,036$), we show that the exponential mixture performance is comparable to existing methods for both small and large datasets. Finally, in simulation studies ($n = 5,036$), we find that the finite exponential mixture exhibits small bias relative to the true statistical uncertainty while maintaining consistent nominal coverage in the bulk.

hep-ex↗

arXiv2026

Mixture-of-Finite-Mixtures Wishart Model for Clustering Covariance Matrices with an Application to Brain Functional Connectivity

Data represented as covariance-type matrices arise in many fields, including brain functional connectivity and diffusion tensor imaging. We develop the MFM-Wishart, a Bayesian model-based clustering approach for such data that combines Wishart mixture components with a mixture-of-finite-mixtures (MFM) prior, allowing joint posterior inference on both the number of clusters and clustering assignments. Theoretically, we study the properties of Wishart kernels in the context of mixture models and then establish results for posterior consistency for the number of clusters and posterior contraction of the mixing measure under standard regularity conditions. Computationally, we develop an efficient Markov chain Monte Carlo (MCMC) algorithm for posterior inference. Simulation studies show competitive clustering performance and accurate recovery of the number of clusters, even under model misspecification. We apply MFM-Wishart to cluster infants based on functional connectivity during sleep, estimated from functional near-infrared spectroscopy (fNIRS) data, illustrating the practical utility of the model and revealing interpretable heterogeneity.

stat.ME↗

arXiv2025

A Physics-Informed Spatiotemporal Deep Learning Framework for Turbulent Systems

Fluid thermodynamics underpins atmospheric dynamics, climate science, industrial applications, and energy systems. However, direct numerical simulations (DNS) of such systems can be computationally prohibitive. To address this, we present a novel physics-informed spatiotemporal surrogate model for Rayleigh-Benard convection (RBC), a canonical example of convective fluid flow. Our approach combines convolutional neural networks, for spatial dimension reduction, with an innovative recurrent architecture, inspired by large language models, to model long-range temporal dynamics. Inference is penalized with respect to the governing partial differential equations to ensure physical interpretability. Since RBC exhibits turbulent behavior, we quantify uncertainty using a conformal prediction framework. This model replicates key physical features of RBC dynamics while significantly reducing computational cost, offering a scalable alternative to DNS for long-term simulations.

physics.flu-dyn↗

arXiv2024

CESAR: A Convolutional Echo State AutoencodeR for High-Resolution Wind Forecasting

An accurate and timely assessment of wind speed and energy output allows an efficient planning and management of this resource on the power grid. Wind energy, especially at high resolution, calls for the development of nonlinear statistical models able to capture complex dependencies in space and time. This work introduces a Convolutional Echo State AutoencodeR (CESAR), a spatio-temporal, neural network-based model which first extracts the spatial features with a deep convolutional autoencoder, and then models their dynamics with an echo state network. We also propose a two-step approach to also allow for computationally affordable inference, while also performing uncertainty quantification. We focus on a high-resolution simulation in Riyadh (Saudi Arabia), an area where wind farm planning is currently ongoing, and show how CESAR is able to provide improved forecasting of wind speed and power for proposed building sites by up to 17% against the best alternative methods.

stat.AP↗

arXiv2024

Modeling High-Resolution Spatio-Temporal Wind with Deep Echo State Networks and Stochastic Partial Differential Equations

In the past decades, clean and renewable energy has gained increasing attention due to a global effort on carbon footprint reduction. In particular, Saudi Arabia is gradually shifting its energy portfolio from an exclusive use of oil to a reliance on renewable energy, and, in particular, wind. Modeling wind for assessing potential energy output in a country as large, geographically diverse and understudied as Saudi Arabia is a challenge which implies highly non-linear dynamic structures in both space and time. To address this, we propose a spatio-temporal model whose spatial information is first reduced via an energy distance-based approach and then its dynamical behavior is informed by a sparse and stochastic recurrent neural network (Echo State Network). Finally, the full spatial data is reconstructed by means of a non-stationary stochastic partial differential equation-based approach. Our model can capture the fine scale wind structure and produce more accurate forecasts of both wind speed and energy in lead times of interest for energy grid management and save annually as much as one million dollar against the closest competitive model.

stat.ML↗

arXiv2024

Boosting Earth System Model Outputs And Saving PetaBytes in their Storage Using Exascale Climate Emulators

We present the design and scalable implementation of an exascale climate emulator for addressing the escalating computational and storage requirements of high-resolution Earth System Model simulations. We utilize the spherical harmonic transform to stochastically model spatio-temporal variations in climate data. This provides tunable spatio-temporal resolution and significantly improves the fidelity and granularity of climate emulation, achieving an ultra-high spatial resolution of 0.034 (approximately 3.5 km) in space. Our emulator, trained on 318 billion hourly temperature data points from a 35-year and 31 billion daily data points from an 83-year global simulation ensemble, generates statistically consistent climate emulations. We extend linear solver software to mixed-precision arithmetic GPUs, applying different precisions within a single solver to adapt to different correlation strengths. The PaRSEC runtime system supports efficient parallel matrix operations by optimizing the dynamic balance between computation, communication, and memory requirements. Our BLAS3-rich code is optimized for systems equipped with four different families and generations of GPUs, scaling well to achieve 0.976 EFlop/s on 9,025 nodes (36,100 AMD MI250X multichip module (MCM) GPUs) of Frontier (nearly full system), 0.739 EFlop/s on 1,936 nodes (7,744 Grace-Hopper Superchips (GH200)) of Alps, 0.243 EFlop/s on 1,024 nodes (4,096 A100 GPUs) of Leonardo, and 0.375 EFlop/s on 3,072 nodes (18,432 V100 GPUs) of Summit.

stat.CO↗

arXiv2023

Physics-Informed Priors with Application to Boundary Layer Velocity

One of the most popular recent areas of machine learning predicates the use of neural networks augmented by information about the underlying process in the form of Partial Differential Equations (PDEs). These physics-informed neural networks are obtained by penalizing the inference with a PDE, and have been cast as a minimization problem currently lacking a formal approach to quantify the uncertainty. In this work, we propose a novel model-based framework which regards the PDE as a prior information of a deep Bayesian neural network. The prior is calibrated without data to resemble the PDE solution in the prior mean, while our degree in confidence on the PDE with respect to the data is expressed in terms of the prior variance. The information embedded in the PDE is then propagated to the posterior yielding physics-informed forecasts with uncertainty quantification. We apply our approach to a simulated viscous fluid and to experimentally-obtained turbulent boundary layer velocity in a water tunnel using an appropriately simplified Navier-Stokes equation. Our approach requires very few observations to produce physically-consistent forecasts as opposed to non-physical forecasts stemming from non-informed priors, thereby allowing forecasting complex systems where some amount of data as well as some contextual knowledge is available.

physics.flu-dyn↗

arXiv2023

A Physics-Informed, Deep Double Reservoir Network for Forecasting Boundary Layer Velocity

When a fluid flows over a solid surface, it creates a thin boundary layer where the flow velocity is influenced by the surface through viscosity, and can transition from laminar to turbulent at sufficiently high speeds. Understanding and forecasting the fluid dynamics under these conditions is one of the most challenging scientific problems in fluid dynamics. It is therefore of high interest to formulate models able to capture the nonlinear spatio-temporal velocity structure as well as produce forecasts in a computationally efficient manner. Traditional statistical approaches are limited in their ability to produce timely forecasts of complex, nonlinear spatio-temporal structures which are at the same time able to incorporate the underlying flow physics. In this work, we propose a model to accurately forecast boundary layer velocities with a deep double reservoir computing network which is capable of capturing the complex, nonlinear dynamics of the boundary layer while at the same time incorporating physical constraints via a penalty obtained by a Partial Differential Equation (PDE). Simulation studies on a one-dimensional viscous fluid demonstrate how the proposed model is able to produce accurate forecasts while simultaneously accounting for energy loss. The application focuses on boundary layer data in a water tunnel with a PDE penalty derived from an appropriate simplification of the Navier-Stokes equations, showing improved forecasting by the proposed approach in terms of mass conservation and variability of velocity fluctuation against non-physics-informed methods.

physics.flu-dyn↗

arXiv2023

A Neural Network-Based Approach to Normality Testing for Dependent Data

There is a wide availability of methods for testing normality under the assumption of independent and identically distributed data. When data are dependent in space and/or time, however, assessing and testing the marginal behavior is considerably more challenging, as the marginal behavior is impacted by the degree of dependence. We propose a new approach to assess normality for dependent data by non-linearly incorporating existing statistics from normality tests as well as sample moments such as skewness and kurtosis through a neural network. We calibrate (deep) neural networks by simulated normal and non-normal data with a wide range of dependence structures and we determine the probability of rejecting the null hypothesis. We compare several approaches for normality tests and demonstrate the superiority of our method in terms of statistical power through an extensive simulation study. A real world application to global temperature data further demonstrates how the degree of spatio-temporal aggregation affects the marginal normality in the data.

stat.ME↗

arXiv2023

Calibrated Forecasts of Quasi-Periodic Climate Processes with Deep Echo State Networks and Penalized Quantile Regression

Among the most relevant processes in the Earth system for human habitability are quasi-periodic, ocean-driven multi-year events whose dynamics are currently incompletely characterized by physical models, and hence poorly predictable. This work aims at showing how 1) data-driven, stochastic machine learning approaches provide an affordable yet flexible means to forecast these processes; 2) the associated uncertainty can be properly calibrated with fast ensemble-based approaches. While the methodology introduced and discussed in this work pertains to synoptic scale events, the principle of augmenting incomplete or highly sensitive physical systems with data-driven models to improve predictability is far more general and can be extended to environmental problems of any scale in time or space.

physics.ao-ph↗

arXiv2023

High Resolution Global Precipitation Downscaling with Latent Gaussian Models and Nonstationary SPDE Structure

Obtaining high-resolution maps of precipitation data can provide key insights to stakeholders to assess a sustainable access to water resources at urban scale. Mapping a nonstationary, sparse process such as precipitation at very high spatial resolution requires the interpolation of global datasets at the location where ground stations are available with statistical models able to capture complex non-Gaussian global space-time dependence structures. In this work, we propose a new approach based on capturing the spatial dependence of a latent Gaussian process via a locally deformed Stochastic Partial Differential Equation (SPDE) with a buffer allowing for a different spatial structure across land and sea. The finite volume approximation of the SPDE, coupled with Integrated Nested Laplace Approximation ensures feasible Bayesian inference for tens of millions of observations. The simulation studies showcase the improved predictability of the proposed approach against stationary and no-buffer alternatives. The proposed approach is then used to yield high resolution simulations of daily precipitation across the United States.

stat.AP↗

arXiv2022

Sensitivity Analysis of Wind Energy Resources with Bayesian non-Gaussian and nonstationary Functional ANOVA

The transition from non-renewable to renewable energies represents a global societal challenge, and developing a sustainable energy portfolio is an especially daunting task for developing countries where little to no information is available regarding the abundance of renewable resources such as wind. Weather model simulations are key to obtain such information when observational data are scarce and sparse over a country as large and geographically diverse as Saudi Arabia. However, output from such models is uncertain, as it depends on inputs such as the parametrization of the physical processes and the spatial resolution of the simulated domain. In such situations, a sensitivity analysis must be performed and the input may have a spatially heterogeneous influence of wind. In this work, we propose a latent Gaussian functional analysis of variance (ANOVA) model that relies on a nonstationary Gaussian Markov random field approximation of a continuous latent process. The proposed approach is able to capture the local sensitivity of Gaussian and non-Gaussian wind characteristics such as speed and threshold exceedances over a large simulation domain, and a continuous underlying process also allows us to assess the effect of different spatial resolutions. Our results indicate that (1) the non-local planetary boundary layer scheme and high spatial resolution are both instrumental in capturing wind speed and energy (especially over complex mountainous terrain), and (2) the impact of planetary boundary layer scheme and resolution on Saudi Arabia's planned wind farms is small (at most 1.4%). Thus, our results lend support for the construction of these wind farms in the next decade.

stat.AP↗

arXiv2022

Calibration of Spatio-Temporal Forecasts from Citizen Science Urban Air Pollution Data with Sparse Recurrent Neural Networks

With their continued increase in coverage and quality, data collected from personal air quality monitors has become an increasingly valuable tool to complement existing public health monitoring systems over urban areas. However, the potential of using such `citizen science data' for automatic early warning systems is hampered by the lack of models able to capture the high resolution, nonlinear spatio-temporal features stemming from local emission sources such as traffic, residential heating and commercial activities. In this work, we propose a machine learning approach to forecast high frequency spatial fields which has two distinctive advantages from standard neural network methods in time: 1) sparsity of the neural network via a spike-and-slab prior, and 2) a small parametric space. The introduction of stochastic neural networks generates additional uncertainty, and in this work we propose a fast approach for ensure that the forecast is correctly assessed (calibration), both marginally and spatially. We focus on assessing exposure to urban air pollution in San Francisco, and our results suggest an improvement of over 58% in the mean squared error over standard time series approach with a calibrated forecast for up to 5 days.

stat.AP↗

arXiv2021

Forecasting High-Frequency Spatio-Temporal Wind Power with Dimensionally Reduced Echo State Networks

Fast and accurate hourly forecasts of wind speed and power are crucial in quantifying and planning the energy budget in the electric grid. Modeling wind at a high resolution brings forth considerable challenges given its turbulent and highly nonlinear dynamics. In developing countries, where wind farms over a large domain are currently under construction or consideration, this is even more challenging given the necessity of modeling wind over space as well. In this work, we propose a machine learning approach to model the nonlinear hourly wind dynamics in Saudi Arabia with a domain-specific choice of knots to reduce the spatial dimensionality. Our results show that for locations highlighted as wind abundant by a previous work, our approach results in an 11% improvement in the two-hour-ahead forecasted power against operational standards in the wind energy sector, yielding a saving of nearly one million US dollars over a year under current market prices in Saudi Arabia.

stat.AP↗

arXiv2021

Assessing the Reliability of Wind Power Operations under a Changing Climate with a Non-Gaussian Bias Correction

Facing increasing societal and economic pressure, many countries have established strategies to develop renewable energy portfolios, whose penetration in the market can alleviate the dependence on fossil fuels. In the case of wind, there is a fundamental question related to the resilience, and hence profitability of future wind farms to a changing climate, given that current wind turbines have lifespans of up to thirty years. In this work, we develop a new non-Gaussian method data to simulations and to estimate future wind, predicated on a trans-Gaussian transformation and a cluster-wise minimization of the Kullback-Leibler divergence. Future winds abundance will be determined for Saudi Arabia, a country with a recently established plan to develop a portfolio of up to 16 GW of wind energy. Further, we estimate the change in profits over future decades using additional high-resolution simulations, an improved method for vertical wind extrapolation, and power curves from a collection of popular wind turbines. We find an overall increase in the daily profit of $272,000 for the wind energy market for the optimal locations for wind farming in the country.

stat.AP↗

arXiv2020

Improving Bayesian Local Spatial Models in Large Data Sets

Environmental processes resolved at a sufficiently small scale in space and time will inevitably display non-stationary behavior. Such processes are both challenging to model and computationally expensive when the data size is large. Instead of modeling the global non-stationarity explicitly, local models can be applied to disjoint regions of the domain. The choice of the size of these regions is dictated by a bias-variance trade-off; large regions will have smaller variance and larger bias, whereas small regions will have higher variance and smaller bias. From both the modeling and computational point of view, small regions are preferable to better accommodate the non-stationarity. However, in practice, large regions are necessary to control the variance. We propose a novel Bayesian three-step approach that allows for smaller regions without compromising the increase of the variance that would follow. We are able to propagate the uncertainty from one step to the next without issues caused by reusing the data. The improvement in inference also results in improved prediction, as our simulated example shows. We illustrate this new approach on a data set of simulated high-resolution wind speed data over Saudi Arabia.

stat.AP↗

arXiv2019

A Non-Gaussian Spatio-Temporal Model for Daily Wind Speeds Based on a Multivariate Skew-t Distribution

Facing increasing domestic energy consumption from population growth and industrialization, Saudi Arabia is aiming to reduce its reliance on fossil fuels and to broaden its energy mix by expanding investment in renewable energy sources, including wind energy. A preliminary task in the development of wind energy infrastructure is the assessment of wind energy potential, a key aspect of which is the characterization of its spatio-temporal behavior. In this study we examine the impact of internal climate variability on seasonal wind power density fluctuations over Saudi Arabia using 30 simulations from the Large Ensemble Project (LENS) developed at the National Center for Atmospheric Research. Furthermore, a spatio-temporal model for daily wind speed is proposed with neighbor-based cross-temporal dependence, and a multivariate skew-t distribution to capture the spatial patterns of higher order moments. The model can be used to generate synthetic time series over the entire spatial domain that adequately reproduce the internal variability of the LENS dataset.

stat.AP↗

arXiv2018

Marginally Parametrized Spatio-Temporal Models and Stepwise Maximum Likelihood Estimation

In order to learn the complex features of large spatio-temporal data, models with large parameter sets are often required. However, estimating a large number of parameters is often infeasible due to the computational and memory costs of maximum likelihood estimation (MLE). We introduce the class of marginally parametrized (MP) models, where inference can be performed efficiently with a sequence of marginal (estimated) likelihood functions via stepwise maximum likelihood estimation (SMLE). We provide the conditions under which the stepwise estimators are consistent, and we prove that this class of models includes the diagonal vector autoregressive moving average model. We demonstrate that the parameters of this model can be obtained at least three orders of magnitude faster using SMLE compared to MLE, with only a small loss in statistical efficiency. We apply an MP model to a spatio-temporal global climate data set (in order to learn complex features of interest to climate scientists) consisting of over five million data points, and we demonstrate how estimation can be performed in less than an hour on a laptop.

stat.CO↗