ELLIS Machine Learning Insights Seminar – Giovanni Catalani (Airbus) and Cristian Meo (TU Delft)
26 November 2024 15:00 till 17:00 - Location: ECHO, Hall B2 - By: ELLIS Delft | Add to my calendar
Information and Schedule
The ELLIS Unit Delft will host a Machine Learning Insights Seminar on 26 November featuring two presentations:
15:00: ‘Autoregressive Deep State Space Models’ by Cristian Meo (PhD Candidate, TU Delft)
15:45: Break (with coffee and tea)
16:00: ‘Neural Fields for Physical Simulations’ by Giovanni Catalani (PhD Candidate, Airbus/ ISAE-SUPAERO)
16:45: Closing
Cristian Meo | Computer Science at TU Delft
Abstract
Autoregressive Deep State Space Models are a powerful tool for modeling temporal dynamics in intelligent systems, combining the capabilities of VQ-VAE, Transformers, and autoregressive techniques. In this talk, we will delve into the core components of these models, including their encoder-decoder structures, feature extraction methodologies, and dynamics modules. We will discuss the applications of these models in video generation, model-based Reinforcement Learning and extreme precipitation forecasting, highlighting their ability to improve predictive accuracy and sample efficiency in real-world scenarios. Additionally, we will address the limitations related to computational constraints and the interpretability of learned representations. This presentation aims to provide insights into how these models are pushing the boundaries of predictive modeling, particularly in decision-making systems requiring temporal reasoning.
Speaker Biography
Cristian Meo (cmeo97.github.io) is a PhD candidate in Computer Science at TU Delft, supervised by Prof. Justin Dauwels and Prof. Geert Leus and mentored by Anirudh Goyal, Research Scientist at Google DeepMind. His research focuses on Generative Modeling, Unsupervised Representation Learning, and Model-Based Reinforcement Learning (RL). Currently, Cristian is working on leveraging pretrained Vision-Language Models (VLMs) to develop Generalist World Models, with a particular interest in generative models for video prediction and model-based RL downstream tasks. Cristian holds a Bachelor's degree in Mechanical Engineering from Politecnico di Torino and a Master's degree in BioRobotics from TU Delft, where he graduated cum laude. During his academic journey, he completed a research internship at Mila – Quebec AI Institute, under the supervision of Prof. Yoshua Bengio (Turing Award 2018), contributing to projects on Hierarchical World Models, disentanglement, and object-centric world models.
Giovanni Catalani | Airbus / École Nationale de l'Aéronautique et de l'Espace (ISAE-SUPAERO), Toulouse, France
Abstract
Recent advancements in deep learning offer powerful tools for accelerating simulations of complex systems governed by Partial Differential Equations (PDEs), particularly in computational fluid dynamics (CFD), structural mechanics, and climate modeling. Traditional numerical solvers, while highly accurate, can be computationally expensive for preliminary analysis or optimization where an extensive exploration of the design space is performed. On the other hand, Neural Fields offer a powerful framework for Operator Learning by parameterizing continuous functions over the physical space using neural networks, enabling discretization-invariant representations of arbitrary functions. In this talk, I present the foundations of Neural Operator learning for the development of scalable, data-driven approaches for large-scale simulations. Moreover, I explore the applications of these methods to aerodynamic simulations for aircraft design. Specifically, I showcase how Neural Fields can be used to construct real-time fluid dynamics simulators over aircraft geometries with shape variations across a wide range of flight conditions. This approach opens up new possibilities for efficient exploration of design spaces, allowing rapid iterations in the design and optimization process. Ultimately, these methods represent a powerful framework for the design and analysis of next-generation aircraft, significantly accelerating development cycles and improving overall efficiency.
Speaker Biography
Giovanni Catalani is a PhD student at Airbus and at the École Nationale de l'Aéronautique et de l'Espace (ISAE-SUPAERO) in Toulouse, France. My research focuses on the application of Deep Learning to Fluid Dynamics and physical simulations of aircraft, aiming to develop scalable, data-driven models for aerodynamic analysis and design optimization. Prior to this, I completed my Master’s degree in Aerospace Engineering at TU Delft, where my thesis at the Netherlands Aerospace Centre (NLR) focused on data-driven models for predicting unsteady aerodynamics of military aircraft.