99524 - GUEST LECTURES AND SEMINARS ON CLIMATE SCIENCE

Learning outcomes

This course made in the format of seminars and guest lecturers will expose the student to the frontier of knowdledge of climate and will apprenhend what are the topics available for the final thesis. The student will be able to grasp what are the emerging areas on climate science and be able to select the topic for future deepening of the knowldge.

Course contents

The course is structured with 1- or 2-hours long time slots and with three types of offers:

1) Seminars: >=1 hour on current research/technological challenges, delivered by specialist.

2) Lecture: >=2 hours on a more general topic of broader relevance and less technical details.

3) Short course: >=3 hours on an additional supplementary skill. Examples may include a focus on programming or on an area of transversal interest.

Speakers and teachers will largely come from other institutions including, University of Reading (UK), UCL/RMI (Belgium), CMCC (IT), CNR (IT), University of Cologne (DE) among others.

Speakers and teachers will largely come from other institutions including, University of Reading (UK), UCL/RMI (Belgium), CMCC (IT), CNR (IT), University of Cologne (DE) among others.

The detailed program follows below. It may be progressively updated as long as we get confirmation and details from the speakers.

06 Oct 16.00-18.00

Lecturer: Ali Aydogdu

Ocean Data Assimilation: Practical aspects and case of Mediterranean analysis and reanalysis systems.

Ali Aydogdu, CMCC, Italy. ali.aydogdu@cmcc.it [mailto:ali.aydogdu@cmcc.it]

Ocean data assimilation (DA) has several aspects differing from other DA applications such as in the atmosphere, due to for example coastal boundaries, relatively slower time variability, less number and limited coverage of observations, model resolution. In this seminar, we will present state-of-the-art data assimilation (DA) techniques commonly used in ocean DA. Some differences with other DA applications in Earth Sciences will be highlighted. Some considerations that are differing from Global Ocean to regional seas to coastal scales will be discussed while available and commonly used observation types will be introduced. Examples from various ocean data assimilation systems either using Kalman filter based schemes or variational schemes be presented with a focus on the Mediterranean Sea analysis and reanalysis systems, a component of Copernicus Marine Service, based on a 3D ocean variational data assimilation scheme OceanVar, developed and maintained at CMCC.

20 Oct 16.00-18.00

Lecturers: Thanh T.N. Nguyen & Hung An Nguyen

Improving air pollution monitoring and management of Vietnam with satellite PM2.5 observation

Thanh T.N. Nguyen, VNU University of Engineering and Technology, thanhntn@vnu.edu.vn [mailto:thanhntn@vnu.edu.vn]

Air pollution is considered an urgent problem in Vietnam as a result of remarkable economic development and urbanization as well as effects of regional transportation. Among air pollution problems in Vietnam, fine Particulate Matter (PM_2.5) pollution is currently one of the most alarming, with multiple negative impacts on public health. In Vietnam, PM_2.5 monitoring has been conducted by local and national governmental organizations, and others such as the US embassy, French Embassy, etc. Despite the immense effort put into monitoring air quality, both the government-controlled and the private-sector company monitoring systems are still inadequate in giving a complete picture of air pollution at a national scale. Furthermore, monitoring data is still limited in distribution, meanwhile air pollution reports remain sparsely published. On the other hand, satellite technology is considered a more modern and expansive approach to providing daily air pollution monitoring solutions from global to local scale. In this presentation, we proposed the methodology for supplementing PM_2.5 monitoring at a national scale based on remote sensing approach and embedded research translation products developed from research results to provide insight into the status of PM_2.5 concentration across the country to communities.

Precipitation estimation using machine learning and deep learning techniques from Himawari-8 satellite imagery and ancillary datasets

Hung An Nguyen⁽¹⁾, Truong X. Ngo⁽²⁾

⁽¹⁾ Le Quy Don Technical University, Hanoi, Vietnam, hungan@lqdtu.edu.vn [mailto:hungan@lqdtu.edu.vn]

⁽²⁾ VNU University of Engineering and Technology, Hanoi, Vietnam, truonggnx@vnu.edu.vn [mailto:truonggnx@vnu.edu.vn]

Rain is a common natural weather phenomenon and plays a crucial role in the hydrological cycle. Changes in rainfall patterns are one of the major factors directly contributing to phenomena such as droughts, floods, climate variability in different regions, as well as causing erosion and altering soil fertility. Rainfall is a meteorological parameter that exhibits continuous spatial and temporal variation. Therefore, accurately estimating rainfall to mitigate the damage it can cause has been and continues to be a significant concern for scientific researchers.

In practice, various devices and methods are employed to estimate rainfall, including ground-based rain gauge stations, weather radar, and meteorological satellites, along with diverse sources of data. Rainfall data measured at gauge stations has good quality but limited spatial coverage (point measurements). Radar data has better spatial coverage (maps) and relatively good quality but comes with higher operational and maintenance costs. Satellites provide data with even broader coverage but lower quality. Research is being conducted to develop methods that combine different data sources for rainfall estimation.

This study proposes a machine learning solution (Random Forest) and a deep learning model (UNET model) using Himawari-8 satellite imagery and additional data sources (ERA-5 meteorological data and ASTERDEM) to enhance the accuracy of rainfall estimation from satellite data. Our research scope covers the North Central region of Vietnam during the two-year period of 2019 and 2020.

26 Oct 17.00-18.00

Lecturer: Yumeng Chen (U of Reading, UK)

Application of data assimilation to neXtSIM: a sea ice model using novel Maxwell-Elasto-Brittle rheology

Advanced data assimilation (DA) methods, widely used in geophysical and climate studies to merge observations with numerical models, can improve the state estimates and consequent forecasts. In recent years, sea ice models using novel Maxwell-Elasto-Brittle (MEB) or Brittle Bingham-Maxwell (BBM) rheology are developed for operational sea ice forecast. These novel sea ice models can have a better representation of the sea ice properties. In this seminar, we will present the state and parameter estimation of the sea ice model using DA method. The state estimation is performed for the sea ice model, neXtSIM. The model is solved on a time-dependent evolving mesh making it a challenging application for ensemble DA. As a solution, we perform the DEnKF analysis on a fixed and regular reference mesh, where model variables are interpolated before the DA and then back to each member's mesh after the DA. We evaluate the impact of assimilating different types of sea-ice observations on the model's forecast skills of the Arctic sea ice by comparing satellite observations and a free-run ensemble in an Arctic winter period, 2019-2020. We then investigate the parameter estimation using a dynamics-only MEB sea ice model on a fixed Eulerian grid. Using idealised experiments, we investigate the possibility of estimating un-observed model variables and model parameters such as the air drag coefficient and damage parameter of the sea ice model. Mimicking the realistic observation network with different combinations of observations, we demonstrate that various issues can potentially arise in a complex sea ice model. We also show that, with the current observation network, it is possible to improve both the observed and unobserved model state forecast and parameters accuracy.

27 Oct 16.00-18.00

Lecturer: James Renwick

10 Nov 16.00-18.00

Lecturer: Valerio Lembo (CNR, Italy)

From the linear response theory to tipping points. How do we investigate the future climate response to anthropogenic GHG forcing?

Projecting the long-term response of the climate system to an external forcing is a complicated endeavor, for which we need to resort either to huge computational resources or advanced mathematical tools (or both). Here, inspired by 2021 Nobel prize winner Klaus Hasselmann’s seminal works, I will try to guide you through the different degrees of complexity that make up our understanding of the climate response, and to what extent the natural variability comes into play. I will address the Hasselmann’s model from the point of view of the cold start problem, then I will briefly point at the optimal fingerprinting methodology for the attribution of climate change to external forcings. Commenting on the feasibility of the approach separating oceanic and atmospheric scales of variability, I will provide some examples of what can be achieved with the linear response theory in terms of climate projections and limitations thereof. Finally, I will focus on when nonlinearity kicks, through the emergence of tipping points, climate surprises, Melancholia states, and how the research community tries to cope with these strange objects hindering our efforts of attaining a smooth docile predictable modeled climate.

17 Nov 16.00-18.00

Lecturer: Ivo Pasmans (U of Reading, UK)

Computational challenges in operational data assimilation: problems and solutions

Operational weather and ocean forecasting proceeds as a sequence of time intervals. During each interval numerical models produce a forecast, observations are collected and a comparison between the two is made. This comparison is used, in a process called data assimilation (DA), to construct observation-informed initial conditions for the forecast in the next time interval. Many DA algorithms are in use, but they all share the need to solve a high-dimensional (>10¹⁰) system of linear equations. Constructing and solving this system in the limited amount of time available between the reception of the observations and the start of the next time interval is highly non-trivial for three reasons. 1) As the numerical models are computationally demanding, it is generally impossible to construct the full linear system. 2) Its high-dimensionality makes it impossible to store the system as a matrix in memory. Consequently, it is not possible to directly invert it. 3) The operational time-constraints strongly limit the number of iterations that can be used by iterative linear solvers. By adapting DA algorithms to use parallelization, it is possible to leverage the computational power of superclusters to construct a high-rank approximation to the linear system and solve it using less then ~20 iterations. In this talk, I will first introduce the two most popular families of DA algorithms: Kalman filters and variational DA. After this, I will discuss some of the adaptations that have been developed to enable parallelization. Among these are ensemble Kalman filters, domain localization, the EVIL (Ensemble Variational Integrated Localized) and saddle point algorithms.

23 Nov 17.00-18.00

Lecturer: Ivo Pasmans (U of Reading, UK)

Tailoring Data Assimilation for Models using Discontinuous Galerkin Methods

Discontinuous Galerkin (DG) methods are rapidly gaining popularity in the geophysical community. In these methods the model solution in each grid cell is approximated as a linear combination of basis functions. Ensemble data assimilation (EnDA) aims to bring the model closer to the truth by combining it with observations using error statistics estimated from an ensemble of model runs. It is known to suffer from several well-documented issues. We have tested whether the DG structure can be exploited to address the following three issues: 1) reduce the need for observation thinning, 2) reduce errors in gradients, 3) produce a more accurate localised ensemble covariance. Using an idealised test setup it is found that strong reduction in error can be realised, especially for high DG orders. However, this does not result in reduction of the error in the gradients. The DG basis is found to be expedient for scale-dependent localisation resulting in an ensemble covariance that is closer to the truth than one created using conventional, non-scale dependent localisation.

 

30 Nov 17.00-18.00

Lecturer: Claudia Acquistapace (Institute for geophysics and meteorology, University of Cologne, DE)

Lecture series: Featuring AI methods applied to cloud satellite observations

Total number of hours: 8

Description of the seminar series. This series of lectures focuses on the topic of computer vision since this research branch, already finding huge applications in very diverse fields, has an enormous potential also for advancing in the atmospheric field. We will go through the working principle of convolutional neural networks, exploring the different aspects of the so-called deep learning methods. We will also introduce recent works that exploit deep learning methods to analyze satellite cloud and precipitation observations. We will show how classification methods can be used to characterize cloud fields and how the Long Short Term Memory models can help in the predicting precipitation. The course is presenting very recent research results and it will create an occasion for everyone, students and lecturer as well, to learn more about these topics.

Resources: In preparation, contact the lecturer for additional info. Detailed material and references will be provided on the first lecture and made available online at this page https://www.claudiaacquistapace.it/activities/teaching.html and/or on the website https://expats-ideas4s.com/

Classification of images using a data driven approach: Convolutional neural networks (CNN) explained

In this first lecture, we will start by discussing how to perform the task of assigning a label (from a given ensemble of labels) to an image with a computer. We will introduce a linear classifier (fully connected) based on a score function that maps images to labels and a loss function that can quantify how good is the agreement between the assigned labels and the image truth. Then we will understand how the network is learning with a process called optimization, that includes various processes behind it: stochastic gradient descent, backpropagation etc. Finally, we will look at the neural network architecture and its different layers and how they are spatially arranged, together with an overview of the functions and parameters involved. We will conclude with some methods to visualize convolutional neural networks and corresponding examples.

Keywords: Data driven approach, k-nearest neighbor, classification and optimization tasks with stochastic gradient descent, backpropagation, neural network architecture, activation functions, spatial arrangement, layer sizing patterns, hyperparameters

01 Dec 16.00-18.00

Lecturer: Claudia Acquistapace

Learning and evaluation of a CNN: babysitting the learning process

We will start with a small recap from the previous lecture and we will then dig into how to prepare the data, initialize weights and run the network. We will discuss the batch normalization and will present tips and tricks that reduce the risks of overfitting and improve the network performance, like regularization, L2 dropout and data augmentation. We will then introduce one example, Resnet, which is often used in meteorological applications.

Keywords: Preprocessing, weight initialization, batch normalization, regularization and L2 dropout, loss functions, data augmentation. Overview of some checks to perform for monitoring the CNN algorithm, using one example. Resnet, fine-tuning, transfer learning

07 Dec 17.00-18.00

Lecturer: Claudia Acquistapace

Applying CNN to improve our understanding about clouds and precipitation.

After the first three hours of theory on CNN, it is time to see how this powerful method can contribute to increase our understanding of cloud and precipitation processes. In this lecture we will give a detailed look at recent research works that exploited CNN to classify cloud mesoscale patterns. We will introduce supervised, unsupervised and self-supervised methods and we will describe how they are used in the different research works. Finally, we will also present current research work done by the EXPATS research group and present the main open research questions they are currently working on.

Keywords: supervised, unsupervised, self-supervised learning, human-label

 

14 Dec 15.30-17.00

Lecturer: Claudia Acquistapace

Recurrent neural networks (RNN): LSTM models and their application for nowcasting precipitation

In this lecture we will introduce the usage of recurrent neural networks to model sequences of data. We will talk about the architecture of the RNN, the problems associated with backpropagation and the Long Short Term Memory model, that tries to mitigate the issues that RNN can encounter. We will then conclude the lecture with an example of application of the LSTM model in the prediction of precipitation fields.

Keywords: LSTM, recurrent neural networks, ifog, exploding gradients, vanishing gradients, gradient clipping

15 Dec 16.00-18.00

Lecturer: Francesco Ragone (UCL, BE)

Simulation of Extreme Events in Climate Models with Rare Event Algorithms

Abstract. The study of extreme events is one of the main areas of application of numerical climate models. Events like heat waves, floods or wind storms, have huge impacts on human societies, and a better understanding of their statistics and predictability is crucially important. Studying extreme and rare events on a robust statistical basis with complex climate models is however computationally challenging, as very long simulations and/or very large ensembles are needed to sample a sufficient number of events. This problem can be tackled using rare event algorithms, numerical tools dedicated to reducing the computational effort required to sample rare events in numerical models. These methods typically take the form of genetic algorithms, where a set of suppression and cloning rules is applied to members of an ensemble simulation, in order to oversample trajectories leading to the events of interest. We show recent applications of these methods to the simulation of heat waves and warm summers in the Northern hemisphere. We show how a rare event algorithm allows to efficiently sample events characterised by persistency of high regional surface temperatures, and we analyse the emergence of atmospheric teleconnections during the events. We then present applications to extremes of Arctic sea ice reduction and weakening of the Atlantic Meridional Overturning Circulation, and we discuss how these results open the way to further applications to a wide range of problems.

22 Dec 16.00-18.00

Lecturer: Claudia Acquistapace

From CNN to attention model and vision transformers for image classification tasks

In this last lecture, we will introduce the vision transformers (ViT). ViTs are models that recently outperformed CNN in many computer vision tasks. For years and until 2017, the CNN models represented the most capable model in performing image classification tasks. ViTs are deep learning models that weight the input data in a differential way based on self-attention mechanisms. We will describe their architecture and explain in what they differ from RNNs, trying to understand the implication of such differences. Finally, we will show some of the main computer vision tasks they are able to achieve. If time allows, we will conclude our seminar series with a brief overview of the deep learning methods for video prediction.

Keywords: Vision transformers, self-attention, frame prediction, patches, linear embeddings

29 Feb 16.00-18.00

Lecturer: Tatsiana Bardachova (KU, Germany)

Towards the assimilation of dual-polarization radar data

The forecast accuracy of numerical weather prediction models is strongly determined by the precision of the initial conditions, especially for storm and convective-scale weather prediction. Since radars allow to capture the internal structure and important microphysical and dynamical processes in convective systems, they are crucial instrument for improvement of weather forecasts on these scales. Dual-polarization radar, in contrast to a prevalent single-polarization radar, also provides information on the types and sizes of hydrometeor particles. As a result, polarimetric radar data (PRD) proves to be a valuable data source for data assimilation. However, direct assimilation of PRD is not used in current operational non-hydrostatic convection-permitting numerical models. This is associated with several difficulties, such as model error estimation, which require further study. The current focus of our study is to directly assimilate PRD in an idealized setup. Subsequently, assimilation of PRD will also be conducted in presence of model error using modern methods of stochastic neural network based machine learning for error estimation. For our study, observation system simulation experiments (OSSEs) were performed that simulate the development of a long-lived supercell using the ICON model with two-moment microphysics scheme. The new polarimetric radar forward operator EMVORADO-POL developed at Deutscher Wetterdienst (DWD) was incorporated in the setup. The first steps towards the direct assimilation of differential reflectivity, in addition to non-polarimetric variables, have been implemented and will be presented.

21 Mar 16.00-18.00

Lecturer: Sebastian Reich (University of Potsdam, Germany)

AULA A

Piano Terra

Edificio in Bo - via Irnerio 46

On the Schrödinger bridge approach to sequential data assimilation

Standard computational approaches to sequential data assimilation build upon two steps: first one predicts future states, second one corrects those states once observations have becoming available. A complete data assimilation cycle then propagates the previous probabilistic distribution of states to the current one. Viewed as such, one senses that the task of data assimilation must have close connections to concepts from coupling of measures. Indeed, such a connection to, so called Schrödinger bridge problems, exists. In this talk, this connection will first be explained and then computational challenges are addressed within the context of established data assimilation techniques such as the ensemble Kalman filter.

 

10 Apr 16.00-18.00

Lecturer: Ieuan Higgs (University of Reading, UK)

AULA 4

Piano Primo Interrato
Edificio in Bo - viale Berti Pichat 6-6/2

Complex network and machine learning to improve marine ecosystem modelling and data assimilation

We aim to better understand connectivity both in space and between biogeochemcical variables. These ideas are then used with a combined ensemble-ML approach to outperform the heavily resource-limited marine biogeochemistry operational system for the North West European shelf seas.We use complex network theory to better represent and understand the ecosystem connectivity in a shelf sea environment. The baseline data used for the analysis are obtained from a state-of-the-art coupled marine physics–biogeochemistry model simulating the North West European Shelf (NWES). The complex network built on model outputs is used to identify the functional groups of variables behind the biogeochemistry dynamics, suggesting how to simplify our understanding of the complex web of interactions within the shelf sea ecosystem. We demonstrate that complex networks can also be used to understand spatial ecosystem connectivity, identifying both the (geographically varying) connectivity length-scales and the clusters of spatial locations that are connected. We show that the biogeochemical length-scales vary significantly between variables and are not directly transferable. We also find that the spatial pattern of length-scales is similar across each variable, as long as a specific scaling factor for each variable is taken into account. The clusters indicate geographical regions within which there is a large exchange of information within the ecosystem, while information exchange across the boundaries between these regions is limited. The results of this study describe how information is expected to propagate through the shelf sea ecosystem, and how it can be used in multiple future applications such as stochastic noise modelling, data assimilation, or machine learning.Furthermore, we use machine learning to develop a model that performs a multivariate postprocessing step on the data assimilation scheme implemented in the operational Met Office biogeochemical marine forecast system.Our ML approach aims at imprinting nonlinearity in the DA step, and also to reduce the need for large, expensive ensemble methods to estimate the background statistics required for updating the unobserved variables in the system. The data used for this are obtained from a state-of-the art coupled marine physics-biogeochemistry model (ERSEM) simulating the vertical column of a coastal station (L4) in the English Channel. The model is trained on the output of a large ensemble of free runs with randomly distributed key biogeochemical model parameters. The results of this study show that we can effectively propagate an analysis increment of a single variable in a system where the background statistics can often be poorly estimated. This is a promising step towards using machine learning emulators as a viable alternative to large, expensive ensembles, for an operational setting where the number of ensemble members can often be a limiting factor.

 

 

11 Apr 16.00-18.00

AULA A

Piano Terra

Edificio in Bo - via Irnerio 46

Lecturer: Simon Driscoll (University of Reading, UK)

Emulating melt ponds on sea ice with neural networks

Sea ice, frozen seawater that floats on the ocean surface, plays an essential role in global ocean circulation and in regulating Earth's climate and weather, and melt ponds that form on the ice have a profound impact on the Arctic's climate by altering the ice albedo. Melt pond evolution is complex, sub grid scale and poorly understood - and melt ponds are represented in sea ice models as parametrisations. To understand the uncertainty in these parametrisations, we perform perturbed parameter ensembles (sampled within known ranges of uncertainty) and Sobol sensitivity analysis on the state-of-the-art sea ice column physics model 'Icepack”. These perturbed parameter ensembles show that the Icepack model demonstrates a very large and spatiotemporally varying sensitivity to its melt pond parameters - with predicted sea ice thickness over the Arctic varying by many metres after only a decade. Understanding the sources of uncertainty, improving parametrisations and fine tuning the parameters is a paramount, but usually very complex and difficult task. Given this uncertainty and source of prediction error, we propose to replace the sub grid scale melt pond parametrisation (MPP) with a data driven emulator. This work is approached in two steps: first using 'perfect' or 'model' data, we show show that neural networks can learn and replace this MPP, and run in the Icepack model without instability or drift. Secondly, using observational data we develop an emulator to predict the observed melt pond state given observational input of climatological variables. Our work contributes to a broader discussion on how data driven neural networks can replace or compliment the `parametric’ sub grid scale parametrisation approach in climate modelling.

 

16 May 16.00-18.00

AULA A

Piano Terra

Edificio in Bo - via Irnerio 46

Lecturer: Mattia Zaramella

 

17 May 16.00-18.00

AULA B

Piano Terra

Edificio in Bo - via Irnerio 46

Lecturer: Col. Paolo Capizzi (Aeronautica Militare)

The WMO defines the particular characteristics that a National Meteorological Service must satisfy to be recognized as such; the main ones are: to make observations (terrestrial and remote sensing), to process the data collected and to interpret the outputs, to archive data and products and to provide their dissemination. These action, in some way, can be carried out by the private sector too, but being the national reference in the international weather context is a task that only a national service can carry out. Not many know which weather organizations Italy is a member of, the role and tasks of these organizations and why we are part of them. With this presentation it will be given an overview of the international meteorological organizations of which Italy is a member State and the role played to date by the AM Meteorological Service in this context.

 

23 May 16.00-18.00

AULA A

Piano Terra

Edificio in Bo - via Irnerio 46

Lecturer: Giuliano Vitali (Department of Agricultural and Food Sciences, University of Bologna)

Observing the micro-climate with the Internet of Things'

Climate change is an issue involving every aspects of society,
and a number of competencies and skillness together with a high degree of multidisciplarity. The perceptions and concerns on climate change are also very different for stakeholders (e.g. food supply chain vs academicians). In the last decades physicist mostly focused on data-analysis while data collection (satellite sensors, and costly observation devices installed on remote sites) hides what should be a 'basic instinct' o physicsts, 'develop a formal knowledge of nature based on their own measurements'. Physics = nature observation (problems) + measurement (electronics) + formal analysis (math) + logics (programming). Today physics is based on electronic measurements (transducers), which is now offering a large number of off-the-shelf low cost devices (controllers, sensors) making easy the development of home devices of any sort. Such availability is harnessing from several years the IoT (Internet of Things) which is also based on cloud computing and web-service development (math+programming). Some case studies will be presented related to the development of low-power devices for micro-climate observations.

Assessment methods

There will not be assessment, however the attendance of at least 70% is required.

Office hours

See the website of Natale Alberto Carrassi