Modelling the stable, nocturnal boundary layer: different approaches to turbulence closure problem

Il lavoro è dedicato all'analisi fisica e alla modellizzazione dello strato limite atmosferico in condizioni stabili. L'obiettivo principale è quello di migliorare i modelli di parametrizzazione della turbulenza attualmente utilizzati dai modelli meteorologici a grande scala. Questi modelli di parametrizzazione della turbolenza consistono nell' esprimere gli stress di Reynolds come funzioni dei campi medi (componenti orizzontali della velocità e temperatura potenziale) usando delle chiusure. La maggior parte delle chiusure sono state sviluppate per i casi quasi-neutrali, e la difficoltà è trattare l'effetto della stabilità in modo rigoroso. Studieremo in dettaglio due differenti modelli di chiusura della turbolenza per lo strato limite stabile basati su assunzioni diverse: uno schema TKE-l (Mellor-Yamada,1982), che è usato nel modello di previsione BOLAM (Bologna Limited Area Model), e uno schema sviluppato recentemente da Mauritsen et al. (2007). Le assunzioni delle chiusure dei due schemi sono analizzate con dati sperimentali provenienti dalla torre di Cabauw in Olanda e dal sito CIBA in Spagna. Questi schemi di parametrizzazione della turbolenza sono quindi inseriti all'interno di un modello colonnare dello strato limite atmosferico, per testare le loro predizioni senza influenze esterne. Il confronto tra i differenti schemi è effettuato su un caso ben documentato in letteratura, il "GABLS1". Per confermare la validità delle predizioni, un dataset tridimensionale è creato simulando lo stesso caso GABLS1 con una Large Eddy Simulation. ARPS (Advanced Regional Prediction System) è stato usato per questo scopo. La stratificazione stabile vincola il passo di griglia, poichè la LES deve essere ad una risoluzione abbastanza elevata affinchè le tipiche scale verticali di moto siano correttamente risolte. Il confronto di questo dataset tridimensionale con le predizioni degli schemi turbolenti permettono di proporre un insieme di nuove chiusure atte a migliorare il modello di turbolenza di BOLAM. Il lavoro è stato compiuto all' ISAC-CNR di Bologna e al LEGI di Grenoble.

Turbulence models in the atmospheric boundary layer under convective conditions

In this work a modelization of the turbulence in the atmospheric boundary layer, under convective condition, is made. For this aim, the equations that describe the atmospheric motion are expressed through Reynolds averages and, then, they need closures. This work consists in modifying the TKE-l closure used in the BOLAM (Bologna Limited Area Model) forecast model. In particular, the single column model extracted from BOLAM is used, which is modified to obtain other three different closure schemes: a non-local term is added to the flux- gradient relations used to close the second order moments present in the evolution equation of the turbulent kinetic energy, so that the flux-gradient relations become more suitable for simulating an unstable boundary layer. Furthermore, a comparison among the results obtained from the single column model, the ones obtained from the three new schemes and the observations provided by the known case in literature ”GABLS2” is made.

Exploiting the clustering of cosmic voids as a novel cosmological probe

The investigations of the large-scale structure of our Universe provide us with extremely powerful tools to shed light on some of the open issues of the currently accepted Standard Cosmological Model. Until recently, constraining the cosmological parameters from cosmic voids was almost infeasible, because the amount of data in void catalogues was not enough to ensure statistically relevant samples. The increasingly wide and deep fields in present and upcoming surveys have made the cosmic voids become promising probes, despite the fact that we are not yet provided with a unique and generally accepted definition for them. In this Thesis we address the two-point statistics of cosmic voids, in the very first attempt to model its features with cosmological purposes. To this end, we implement an improved version of the void power spectrum presented by Chan et al. (2014). We have been able to build up an exceptionally robust method to tackle with the void clustering statistics, by proposing a functional form that is entirely based on first principles. We extract our data from a suite of high-resolution N-body simulations both in the LCDM and alternative modified gravity scenarios. To accurately compare the data to the theory, we calibrate the model by accounting for a free parameter in the void radius that enters the theory of void exclusion. We then constrain the cosmological parameters by means of a Bayesian analysis. As far as the modified gravity effects are limited, our model is a reliable method to constrain the main LCDM parameters. By contrast, it cannot be used to model the void clustering in the presence of stronger modification of gravity. In future works, we will further develop our analysis on the void clustering statistics, by testing our model on large and high-resolution simulations and on real data, also addressing the void clustering in the halo distribution. Finally, we also plan to combine these constraints with those of other cosmological probes.

Meta-learning for informed model searches in Automated Machine Learning

Day by day, machine learning is changing our lives in ways we could not have imagined just 5 years ago. ML expertise is more and more requested and needed, though just a limited number of ML engineers are available on the job market, and their knowledge is always limited by an inherent characteristic of theirs: they are humans. This thesis explores the possibilities offered by meta-learning, a new field in ML that takes learning a level higher: models are trained on other models' training data, starting from features of the dataset they were trained on, inference times, obtained performances, to try to understand the relationship between a good model and the way it was obtained. The so-called metamodel was trained on data collected by OpenML, the largest ML metadata platform that's publicly available today. Datasets were analyzed to obtain meta-features that describe them, which were then tied to model performances in a regression task. The obtained metamodel predicts the expected performances of a given model type (e.g., a random forest) on a given ML task (e.g., classification on the UCI census dataset). This research was then integrated into a custom-made AutoML framework, to show how meta-learning is not an end in itself, but it can be used to further progress our ML research. Encoding ML engineering expertise in a model allows better, faster, and more impactful ML applications across the whole world, while reducing the cost that is inevitably tied to human engineers.