iNEST DAY - February 21, 2025

Registration opens at 8:30

9:00 - 9:10

Mattia Fogagnolo, Matteo Zavatteri

Opening

Overview of the results of the iNEST Spoke 9

Chair: Mario Putti

10:30 - 11:00

Coffee break

11:00 - 13:00

Round table on the role and function of interconnected research ecosystems

Participants:

Moderators: Mattia Fogagnolo, Matteo Zavatteri

13:00 - 14:30

Lunch break*

RT1 scientific session on Mathematical, Numerical, and Data-driven Modeling

Abstract: Calderon’s inverse problem consists in determining the value of the weight in model  governed by a weighted laplacian PDE from boundary measurements only, that mathematically are represented by the Dirichlet-to-Neumann map. It has been recently observed that the p(x)-laplacian version of this problem emerges when trying to detect losses in water distribution systems. In this talk, we will describe a simplified mathematical framework for this problem, highlight the formidable challenges and present some new results


Abstract: Certifying Artificial Intelligence systems has become more and more important, especially when such systems are classified as high-risk. In this talk, we sum up the current results on the automated synthesis of certified neural networks obtained in the contest of iNEST and we discuss challenges, current limitations, open research lines and next steps.

Chair: Mario Putti

RT2 scientific session on Model Order Reduction

Abstract: Turbulence modeling and simulation have long been fundamental in describing a wide range of physical phenomena, from classical aeronautical applications to the dynamics of the entire atmosphere. While significant progress has been made over the years, today's challenge clearly lies in integrating machine learning techniques and, more broadly, leveraging the rapidly growing abundance of data to enhance turbulence modeling, understanding, and simulation. This talk will highlight key areas where machine learning can contribute to turbulence research, including the modeling of coarse-grained solutions, wall modeling, and reduced-order modeling.


Abstract: This talk presents the development of a Digital Shadow of the Mediterranean Copernicus Analysis and Forecast System, a project under the iNEST initiative, jointly developed by SISSA and OGS. A Digital Shadow is a surrogate model designed to replicate the system's behavior efficiently, incorporating data assimilation and uncertainty quantification. The system, consisting of physics and biogeochemistry components, is aimed at improving short-term and long-term forecasts for the Mediterranean region. The methodology is novel and combines three techniques: quadratic manifold dimensionality reduction, Gaussian Process Regression, and Stochastic Differential Equations (SDEs). These methods are employed to approximate system behavior with reduced computational complexity, enabling faster predictions without compromising too much accuracy. The model's performance is validated through the comparison of forecasted and observed data, ensuring that uncertainty quantification aligns with expected confidence intervals. Additionally, parallel computing is utilized by splitting the Mediterranean Sea into different basins to enhance computational efficiency. This approach offers significant potential for both short-term and long-term forecasting applications. The methodology is compared with respect to state of the art techniques.

Chair: Daniela Tonon

15:50 - 16:10

Coffee break

RT3 scientific session on Automatic Learning for Digital Twins

Abstract: Deep generative models, such as generative adversarial networks (GANs) and score-based diffusion models,  have recently emerged as powerful tools for planning tasks and behaviour synthesis in autonomous systems. Various guidance strategies have been introduced to steer the generative process toward outputs that are more likely to satisfy the planning objectives. These strategies avoid the need for model retraining but do not provide any guarantee that the generated outputs will satisfy the desired planning objectives. To address this limitation, we introduce certified guidance, an approach that modifies a generative model – without retraining it – into a new model guaranteed to satisfy a given specification with probability 1. We focus on Signal Temporal Logic (STL) specifications, which are rich enough to describe non-trivial planning tasks. Our approach leverages neural network verification techniques to systematically explore the generative models' latent spaces, identifying latent regions that are certifiably correct w.r.t. the STL property of interest. We evaluate the effectiveness of our method on four planning benchmarks using GANs and diffusion models. Our results confirm that certified guidance produces generative models that are always correct, unlike existing (non-certified) guidance methods.


Abstract: We introduce Limited Rollout Beam Search (LRBS), a beam search strategy for deep reinforcement learning (DRL) based combinatorial optimization improvement heuristics. Utilizing pre-trained models on the Euclidean Traveling Salesperson Problem, LRBS significantly enhances both in-distribution performance and generalization to larger problem instances, achieving optimality gaps that outperform existing improvement heuristics and narrowing the gap with state-of-the-art constructive methods. We also extend our analysis to two pickup and delivery TSP variants to validate our results. Finally, we employ our search strategy for offline and online adaptation of the pre-trained improvement policy, leading to improved search performance and surpassing recent adaptive methods for constructive heuristics

Chair: Alberto Valese

RT4 scientific session on Applications of Digital Twins in Industry, Medicine, Environmental Sciences, Daily Life

Abstract: In regional ocean models, increasing resolution helps capture small- and mesoscale features that remain undetectable at the coarser resolutions used for larger basins. However, high-resolution numerical simulations are computationally expensive and time-consuming. To address this challenge, deep learning methods for super-resolution of marine variables have been increasingly developed in recent years, inspired by techniques from digital image processing. In this study, we focus on the northern Adriatic Sea, a region of the Mediterranean, and present two applications of a super-resolution algorithm. We discuss the strengths and limitations of this approach in the context of marine environmental modeling.


Abstract: Applications of machine learning to Earth system modelling address different scientific questions, and span methodological approaches, Earth system components, time and spatial scales. We will present a high-level overview of current developments in data-driven Earth modelling, drawing from advancements in weather forecasting and with a focus on ocean modelling. In particular, we present current efforts for the development of a global ocean circulation model targeted at subseasonal-to-seasonal time scales and ensemble superresolution.

Chair: Eleonora Spricigo

17:30

Mattia Fogagnolo, Matteo Zavatteri

Closing


* lunch is not included in the registration