Accepted Posters
Poster 1
Authors and affiliations
Erta Beqiri Brain Physics Laboratory, Department of Clinical Neuroscience, Division of Neurosurgery, University of Cambridge, U.K.
Xuhang Chen Brain Physics Laboratory, Department of Clinical Neuroscience, Division of Neurosurgery, University of Cambridge, U.K.
Saliha Afzaal Brain Physics Laboratory, Department of Clinical Neuroscience, Division of Neurosurgery, University of Cambridge, U.K.
Sho Giersztein Brain Physics Laboratory, Department of Clinical Neuroscience, Division of Neurosurgery, University of Cambridge, U.K.
Ihsane Olakorede Brain Physics Laboratory, Department of Clinical Neuroscience, Division of Neurosurgery, University of Cambridge, U.K.
Stefan Yu Bogli Brain Physics Laboratory, Department of Clinical Neuroscience, Division of Neurosurgery, University of Cambridge, U.K.
Tommaso Rochat Brain Physics Laboratory, Department of Clinical Neuroscience, Division of Neurosurgery, University of Cambridge, U.K.
Cameron Smith Brain Physics Laboratory, Department of Clinical Neuroscience, Division of Neurosurgery, University of Cambridge, U.K.
Masumi Tanaka Gutiez Brain Physics Laboratory, Department of Clinical Neuroscience, Division of Neurosurgery, University of Cambridge, U.K.
Pietro Lio Department of Computer science, University of Cambridge, U.K.
Giovanna Maria Dimitri Department of Information Engineering and Mathematics, University of Siena, Italy
Peter Smielewski Brain Physics Laboratory, Department of Clinical Neuroscience, Division of Neurosurgery, University of Cambridge, U.K.
Title
Improving the reliability of the vascular reactivity index for autoregulation guided management through machine learning modeling
Introduction
The pressure reactivity index (PRx) and derived cerebral perfusion pressure targets can be considered digital biomarkers for traumatic brain injury (TBI) patients with intracranial pressure (ICP) monitoring. One methodological challenge related to PRx is that the calculations presume causality of blood pressure (ABP) changes transmitted to ICP. This assumption is violated during suction events, leading to falsely positive PRx estimates. The derived false representation of physiological patterns would be detrimental when the data are used to train digital twin models for TBI patients. Our goal is to assess the impact of these suction events on PRx values and to develop an automated method that could label these periods as unreliable for PRx assessment.
Methodology
A 3-layers fully convolutional neural network was trained for a classification task (suction/no suction) on a balanced dataset of 3308 (train:validation:test = 60:20:20) 10-seconds ICP segments at 120 Hz. The segments were manually labelled as containing suctions (50%) or not. The segments were extracted from 28 different TBI patients. The classifier was used within ICM+ software to mark data before calculating PRx on the 28 de-identified high-resolution TBI recordings (REC 23/YH/0085). DeltaPRx was calculated as PRx_without_suctions - PRx_with_suctions for each minute-by-minute value across all patients. Summaries of DeltaPRx were analysed based on the percentage of suctions involved (>10%, >20%, >30%, >40%). Additionally, distribution plots were generated to compare the two versions of PRx at varying percentages of suctions.
Results
The total duration of the recordings was a median of 7 days (q1: 5, q3: 8). PRx was available for 84% (71 : 90) of the monitoring time. The comparison of PRx values with and without suction periods revealed differences ranging from -1.8 to 1.05, which are concerning as such variation could fundamentally alter the interpretation of PRx. The percentage of PRx data coverage loss due to suctions spanning >50% of the calculation data buffer ranged from 0 to 2.24%.
Conclusion
Our preliminary results underscore the importance of suction events on PRx assessments. A simple model allows for automated detection and rejection of affected segments improving the robustness of PRx assessments. This tool could help to accurately annotate physiological patterns in the data that are used to train digital twins models for TBI patients.
Presenter
Giovanna Maria Dimitri
Poster 2
Authors and affiliations
Elia Brentarolli University of Verona, Italy
Ashraf Sharifi University of Verona, Italy
Davide Quaglia University of Verona, Italy
Sara Migliorini University of Verona, Italy
Luca Benvenuti Sapienza University of Rome, Italy
Tiziano Villa University of Verona, Italy
Title
Computational methodologies to enable the digital twin in precision farming
Abstract
A digital twin in precision agriculture allows to "close the loop" between the well-known crop monitoring and the "next step that the farmer must manually take to implement the best agronomic strategy.
In this paper we want to present two computational methodologies that enable the creation of a digital twin of a crop in horticulture or floriculture.
It will be shown how the formalism of finite state automata can become the unifying modeling methodology of each plant or homogeneous group of plants and of the pathogens that we want to control.
We will also show the methodology of virtual sensors or soft sensors to fine-grain model the climate inside the greenhouse.
Presenter
Elia Brentarolli
Poster 3
Authors and affiliations
Bruno Guindani Politecnico di Milano, Italy
Livia Lestingi Politecnico di Milano, Italy
Matteo Camilli Politecnico di Milano, Italy
Marcello Maria Bersani Politecnico di Milano, Italy
Title
Capturing Medical Knowledge into a Safe and Trustworthy Digital Twin
Abstract
When managing clinical patients in critical health conditions, it is essential that the decisions made by physicians and nurses are both safe and trustworthy. In this work, we propose a framework to develop and validate an accurate and explainable model designed to assist healthcare providers in caring for patients with respiratory issues. Leveraging data-driven learning, the framework builds a Stochastic Hybrid Automaton (SHA) representing the system, which includes the patient, the provider, and a mechanical ventilator, thereby creating a digital twin of the system. This digital twin encodes the relationships among the system’s elements, incorporates the expertise and intuition of physicians, and provides recommendations for adjusting the ventilator’s parameters to the healthcare provider. The proposed framework ensures safety by aligning the model’s suggestions with established clinical practices and achieves trustworthiness through the high accuracy of the learned model, even under uncertainties associated with human behavior.
Presenter
Bruno Guindani
Poster 4
Authors and affiliations
Gaia Saveri University of Trieste, Italy
Laura Nenzi University of Trieste, Italy
Simone Silvetti University of Trieste, Italy
Luca Bortolussi University of Trieste, Italy
Jan Křetínský Masaryk University, Brno, Czech Republic
Title
stl2vec: Semantic and Interpretable Vector Representation of Temporal Logic
Abstract
Integrating symbolic knowledge and data-driven learning algorithms is a longstanding challenge in Artificial Intelligence. Despite the recognized importance of this task, a notable gap exists due to the discreteness of symbolic representations and the continuous nature of machine-learning computations. One of the desired bridges between these two worlds would be to define semantically grounded vector representation (feature embedding) of logic formulae, thus enabling to perform continuous learning and optimization in the semantic space of formulae. We tackle this goal for knowledge expressed in Signal Temporal Logic (STL) and devise a method to compute continuous embeddings of formulae with several desirable properties: the embedding (i) is finite-dimensional, (ii) faithfully reflects the semantics of the formulae, (iii) does not require any learning but instead is defined from basic principles, (iv) is interpretable. Another significant contribution lies in demonstrating the efficacy of the approach in two tasks: learning model checking, where we predict the probability of requirements being satisfied in stochastic processes; and integrating the embeddings into a neuro-symbolic framework, to constrain the output of a deep-learning generative model to comply to a given logical specification.
Presenter
Simone Silvetti
Poster 5
Authors and affiliations
Francesco Biondani University of Verona, Italy
Luigi Capogrosso University of Verona, Italy
Nicola Dall’Ora University of Verona, Italy
Enrico Fraccaroli University of Verona, Italy
Marco Cristani University of Verona, Italy
Franco Fummi University of Verona, Italy
Title
Human-Centered Digital Twin for Industry 5.0
Abstract
Moving beyond the automation-driven paradigm of Industry 4.0, Industry 5.0 emphasizes human-centric industrial systems where human creativity and instincts complement precise and advanced machines. With this new paradigm, there is a growing need for resource-efficient and user-preferred manufacturing solutions that integrate humans into industrial processes. Unfortunately, methodologies for incorporating human elements into industrial processes remain underdeveloped. In this work, we present the first pipeline for the creation of a human-centered Digital Twin (DT), leveraging Unreal Engine's MetaHuman technology to track worker alertness in real-time. Our findings demonstrate the potential of integrating AI and human-centered design within Industry 5.0 to enhance both worker safety and industrial efficiency.
Presenter
Francesco Biondani
Poster 6
Authors and affiliations
Marco Tiraboschi University of Padova, Italy
Giovanni Nicoli University of Padova, Italy
Roberto Barumerli University of Verona, Italy
Andrea Gulli University of Padova, Italy
Michele Geronazzo University of Padova, Italy
Title
S-TWIN: Developing a Digital Twin for Modeling and Enhancing Sonic Interactions between Children and their Cochlear Implants
Abstract
This work presents a foundational study of a digital twin (DT) for cochlear implants (CIs). The primary goal was to explore the capabilities of a DT in modeling and predicting outcomes within the context of speech therapy rehabilitation. A multidisciplinary team developed a speech therapy rehabilitation decision-making diagram: the basis for data analysis and causal interpretation. The proposed DT integrates this workflow into a statistical model, providing a decision support system for determining whether the CI map requires updating. A second step consisted of modeling the individual temporal hearing status data and having the DT predict when the CI map would be adjusted to optimize rehabilitation. Future developments aim to incorporate more heterogeneous data, enabling a broader predictive scope and enhancing diverse hearing abilities.
Presenter
Andrea Gulli
Poster 7
Authors and affiliations
Silvia Bonfanti Department of Management, Information and Production Engineering, University of Bergamo, Italy
Andrea Bombarda Department of Management, Information and Production Engineering, University of Bergamo, Italy
Angelo Gargantini Department of Management, Information and Production Engineering, University of Bergamo, Italy
Alessandro Colombo Department of Engineering and Applied Sciences, University of Bergamo, Italy
Gionatha Pirola Department of Management, Information and Production Engineering, University of Bergamo, Italy
Title
BREATHE: A Digital twin-based Respiratory System Simulator for Mechanical Ventilator Testing and Training
Abstract
Medical simulation has become a crucial element in training and furthering clinical skills, particularly in mechanical ventilation. Integration between respiratory simulators and virtual mechanical ventilators is still a significant challenge. This research aims to develop an innovative solution, based on the use of digital twins, to overcome these limitations by designing a system that allows the interconnection between a respiratory simulator and a virtual mechanical ventilator, intended for testing ventilators under development. In addition, a graphical user interface is implemented, to ensure effective training in the use of the devices. The approach is based on the digital twin concept, creating a virtual representation of the patient that includes his or her clinical condition and pathology, allowing the evolution of vital parameters to be simulated in response to specific events.
Presenter
Silvia Bonfanti
Poster 8
Authors and affiliations
Chiara Braghin Department of Computer Science "Giovanni Degli Antoni", University of Milan, Italy
Stelvio Cimato Department of Computer Science "Giovanni Degli Antoni", University of Milan, Italy
Andrea Marchesini Department of Computer Science "Giovanni Degli Antoni", University of Milan, Italy
Fabio Palazzesi Department of Computer Science "Giovanni Degli Antoni", University of Milan, Italy
Simone Pesci Department of Computer Science "Giovanni Degli Antoni", University of Milan, Italy
Elvinia Riccobene Department of Computer Science "Giovanni Degli Antoni", University of Milan, Italy
Title
Towards Secure and Interoperable Digital Twins for Healthcare Systems
Abstract
Digital Twins have been recently recognized as promising solutions to address the evolving design needs of Healthcare 4.0 services, which require (a) self-sustaining wireless systems capable of operating with minimal user intervention, and (b) proactive online learning-based systems to perform analysis and provide real-time relevant healthcare responses, enable decision-making, and assessing risks for the patients. Despite their potential, implementing Digital Twins in healthcare poses significant challenges due to the complex, highly distributed and shared natured of the required infrastructure, which becomes an attractive target for hackers and malicious actors. Moreover, healthcare is a highly critical domain where sensitive and personal patient data, if compromised, could lead to severe financial and political repercussions.
While Digital Twins offer a transformative opportunity to redefine healthcare services, their adoption is hindered by several multi-faceted challenges. First, optimizing the underlying infrastructure and architecture is essential for managing what is essentially a multi-tenant big data pipeline. This pipeline must ensure robust access control mechanisms to protect sensitive data while simultaneously providing a secure and trustworthy environment that supports data sharing for enhanced diagnostic and predictive capabilities. Achieving this balance calls for a deep focus on security considerations. In addition, generalizing the Digital Twin approach to make it applicable across various healthcare domains requires standardizing data formats and protocols. Such standardization ensures that solutions are reusable and interoperable, extending their utility beyond specific medical applications.
In the poster, we present open research challenges and propose potential solutions for designing secure, scalable, and efficient architectures of Digital Twins for healthcare systems. Using a lung mechanical ventilator as a case study, we explore key challenges related to security, privacy, trust, and safety in our Digital Twin solution.
The work is supported by the MUR PRIN project G53D23002770006 SAFEST “Trust assurance of Digital Twins for medical cyber-physical systems".
Presenter
Simone Pesci
Poster 9
Authors and affiliations
Alessandro Scagliotti TU Munich and Munich Center for Machine Learning (MCML)
Title
Optimal control of ODEs with dynamics uncertainty
Abstract
In this poster, we focus on problems related to the simultaneous control of ensembles of dynamical systems (ODEs), and we propose as an application the optimal design of pharmacological treatments in medical Oncology.
More in general, this framework arises naturally in several situations in Applied Mathematics, for example when a usual control system (eg related to a physical or biomedical model) depends on parameters affected by uncertainty, or when the Cauchy datum is not available with precision due to measurement errors. In this setting, we typically aim at finding a strategy that should be the same for every system of the ensemble, and that minimizes a proper cost. Namely, the proposed policy should incorporate the uncertainty that affects the system, and typically we seek one that results in a good performance in the most likely scenarios (averaged optimization), or one that guarantees resilience in the least favorable conjuncture (worst-case optimization).
Presenter
Alessandro Scagliotti
Poster 10
Authors and affiliations
Guglielmo Padula SISSA, Trieste, Italy
Stefano Querin OGS, Trieste, Italy
Stefano Salon OGS, Trieste, Italy
Gianluigi Rozza SISSA, Trieste, Italy
Title
A GPR-SDE-based Reduced Order Model for Forecasting and Digital Shadows
Abstract
This poster 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 code of the Digital Shadow is available at https://github.com/guglielmopadula/MedDS.
Presenter
Guglielmo Padula
Poster 11
Authors and affiliations
Raj Narayan Dhara University of Padova, Italy
Mattia Fogagnolo University of Padova, Italy
Title
Boundary determination in the inverse problem for the p(x)-laplacian
Abstract
The inverse problem for the p(x)-laplacian consists in determining both the value of the weight and of the exponent p(x) in a boundary value problem for the weighted p(x)-laplacian with Dirichlet data, assuming the knowledge of a suitable Dirichlet-to-Neumann map. This problem has been recently observed to arise in models for water distribution systems. We show how to recover both the weight and the function p(x) on the boundary of the domain.
Presenter
Raj Narayan Dhara
Poster 12
Authors and affiliations
Laura Rinaldi Department of Mathematics “Tullio Levi-Civita”, Università degli Studi di Padova, Italy
Giulio G. Giusteri Department of Mathematics “Tullio Levi-Civita”, Università degli Studi di Padova, Italy
Fabio Marcuzzi Department of Mathematics “Tullio Levi-Civita”, Università degli Studi di Padova, Italy
Title
Digital twin of bread leavening for energy saving
Abstract
One of the main concepts of the fourth industrial revolution is the digital twin (DT) which helps to build a bridge between the physical and the digital world and becomes a fundamental component to check the status of machines, perform predictions, and make decisions to modify their behavior.
In this poster, we present the construction of our digital twin of bread leavening for energy saving. Firstly, we build a leavening model governed by partial differential equations. Secondly, by noticing that, the DT has to run simultaneously with the real process in an embedded system, we formulate a surrogate model by exploiting VarMiON technology. Finally, we try to rebuild the porous texture of bread during leavening by solving inverse problems thus calibrating the model parameters. To this goal, we apply a strategy that shows the equivalence between voids and fictitious forcing terms. An algorithm of energy estimation completes the project and allows energy saving in real-time.
Presenter
Laura Rinaldi
Poster 13
Authors and affiliations
Ester Bergantin Department of Mathematics, University of Trento, Italy
Federica Caforio University of Graz, Austria
Lucas O. Müller Department of Mathematics, University of Trento, Italy
Christoph M. Augustin University of Graz, Austria
Simone Pezzuto Department of Mathematics, University of Trento, Italy
Title
An Integrated Electromechanical-Circulation Model Of The Heart And Application To Ventricular Dyssynchrony
Abstract
Under physiological conditions, the heart’s ventricles experience rapid electrical activation through the cardiac conduction system, resulting in nearly synchronous mechanical contraction. Cardiac electrical asynchrony occurs due to conduction disorders such as left bundle-branch block (LBBB). Electrically asynchronous activation leads to heterogeneous myocardial contraction, potentially causing reduced pump function. Cardiac resynchronization therapy has emerged as an important treatment to improve pump function in patients with LBBB by restoring ventricular electrical synchrony through ventricular pacing. In this context, computational cardiovascular modelling is increasingly recognized and used as a valuable tool for investigating pathological conditions and evaluating treatment strategies.
The main goals of our research involve modelling the effects of dyssynchrony on regional tissue mechanics and global cardiac pump function as well as investigating the effect of different pacing strategies.
Using patient-specific cardiac geometries, electrical activation maps were calculated using an Eikonal model on a 3D biventricular geometry. The Purkinje network was identified using a probabilistic approach based on non-invasive clinical data such as the standard ECG. The activation times were then averaged within the AHA regions and used as inputs in CircAdapt, a cardiovascular model capable of describing heterogeneous mechanical activation. The CircAdapt model was calibrated to match MRI-derived variables such as end-systolic and end-diastolic volumes. We then compared simulated strain, regional myocardial work, and hemodynamic function for each pacing strategy.
Using this coupled approach, we confirmed the existence of a non-trivial relationship between electrical synchrony, mechanical homogeneity, and hemodynamic function. Furthermore, we identified that the pacing sites associated with the most synchronous electrical activation and homogeneous mechanical work distribution were the right ventricular apical endocardium and left ventricular free wall epicardium, when sequentially paced.
Presenter
Ester Bergantin
Poster 14
Authors and affiliations
Bianca Maria Laudenzi Department of Mathematics, University of Trento, Italy
Lucas Omar Müller Department of Mathematics, University of Trento, Italy
Title
In-silico study to predict in-hospital indicators from wearable-derived signals for cardiovascular and cardiorespiratory disease monitoring
Abstract
Cardiovascular and cardiorespiratory diseases (CVRDs) are the leading cause of death worldwide. The construction of health digital twins for patient monitoring is becoming a fundamental tool to reduce invasive procedures, minimize patient hospitalization, design clinical trials and personalize therapies. The aim of this study is to investigate the feasibility of machine learning-based bio-signals analysis for the monitoring of patients with CVRDs, using a database of in-silico patient data.
In particular, a population of virtual subjects, both healthy and with CVRDs, is created using a comprehensive zero-dimensional global closed-loop model, comprising major elements characterizing cardio-respiratory functions, such as the cardiovascular system, the lungs mechanics, the gas exchange and transport and the main short term regulatory mechanism. For the creation of the virtual database, we explore the sensitivity of clinically relevant model output variables to model parameters. The database of virtual subjects is created by varying the parameters with major impact on the CVRDs-relevant signals.
Our database allows us to train Gaussian process regressors , informed by wearable-acquired data, to predict variables normally acquired only during in-hospital exams. Simulated results provides theoretical assessment of accuracy for predictions of stroke volume, cardiac output, ejection fraction, central venous pressure, and partial pressure of oxygen and carbon dioxide, using wearable-derived indices, i.e. systolic/diastolic blood pressures and heart rate.
Results of this study will allow us to make a-priori studies on feasible strategies for the combination of in-hospital and wearable-acquired data, aiming at early detection and monitoring of cardiovascular and cardio-respiratory pathologies. This framework could be applied to real clinical data, evaluating the benefit of using surrogate models, learned in the in-silico study, as prior for the case of real data.
Presenter
Bianca Maria Laudenzi
Poster 15
Authors and affiliations
Antonia Larese Department of Mathematics “Tullio Levi-Civita”, Università degli Studi di Padova, Italy
Francesco Marchetti Department of Mathematics “Tullio Levi-Civita”, Università degli Studi di Padova, Italy
Mario Putti Department of Agronomy, Food, Natural Resources, Animals and Environment, Università degli Studi di Padova, Italy
Nicola Segala Almaviva Bluebit
Title
Modeling water distribution networks via the p-Laplacian operator
Abstract
Precisely modeling an operational Water Distribution System (WDS) is a non-trivial and intricate task. Indeed, many variables and parameters that were set when designing the WDS are likely changed in time, and performing a complete update is very often unrealistic due to the lack of available data. We propose a surrogate non-linear model that is capable of accurately reproducing the whole water dynamics in the WDS despite being ruled by just two vectors of free parameters, which characterize the underlying graph p-Laplacian operator. Numerical tests certify the good properties of the presented model.
Presenter
Nicola Segala
Poster 16
Authors and affiliations
Federica Adobbati OGS, Trieste, Italy
Lorenzo Bonin University of Trieste, Italy
Gianpiero Cossarini OGS, Trieste, Italy
Valeria Di Biagio OGS, Trieste, Italy
Fabio Giordano OGS, Trieste, Italy and University of Trieste, Italy
Luca Manzoni OGS, Trieste, Italy and University of Trieste, Italy
Stefano Querin OGS, Trieste, Italy
Title
Super-resolution U-nets for coastal ocean modelling
Abstract
Current regional-scale oceanographic operational systems may lack the resolution needed for coastal applications, where fine-scale dynamics such as river outflow and local processes are poorly represented. AI-based techniques for interpolation and fine-scale reconstruction can be applied for studying coastal dynamics. We propose a deep learning method based on a UNet-like architecture for coastal downscaling of 3D physical and biogeochemical variables. Our method is applied to the northern Adriatic Sea, a marginal region of the Mediterranean characterized by strong spatial and temporal variability, where river inputs have a significant impact, especially near the coast. We trained a neural network obtained by composing a U-Net and a Multi-Layer Perceptron on a reanalysis dataset, covering the period from 2006 to 2017, with a horizontal resolution of about 750 m, using as input the regional-scale products of the Marine Copernicus Service for the Mediterranean Sea and the flow-rates of 19 rivers entering the area. We show that our architecture is capable of recovering fine-scale features that are not captured by low-resolution modeling systems.
Presenter
Federica Adobbati
Poster 17
Authors and affiliations
Roberta Di Fonso Aalborg University, Denmark
Remus Teodorescu Aalborg University, Denmark
Carlo Cecati Università degli studi dell'Aquila, Italy
Pallavi Bharadwaj Aalborg University, Denmark
Title
From battery signal analysis to battery digital twin synthesis
Abstract
A digital twin is a digital representation of a unique system of interest built by using models, information, and data that are representative of the selected characteristics, behaviors, and life cycle of the system that we want to model. We present practical steps to build a battery digital twin (BDT) starting from pulsed voltage and current signals. The BDT is based on feedforward neural networks trained to output the dynamic values of a 3RC Equivalent Circuit Model (ECM) of the battery. The effectiveness of the proposed method is verified on the dataset from the prognostics data repository of NASA. The BDT could be used: to produce data signals for testing state of charge (SoC) and state of health (SoH) algorithms; to simulate a battery pack with cells at different SoC and/or ageing states; to develop and test cell balancing algorithms. Moreover, the BDT could be run in parallel with a real battery for monitoring and diagnostic purposes.
Presenter
Roberta Di Fonso
Poster 18
Authors and affiliations
Srajan Goyal Fondazione Bruno Kessler, Trento, Italy
Alberto Griggio Fondazione Bruno Kessler, Trento, Italy
Stefano Tonetta Fondazione Bruno Kessler, Trento, Italy
Title
Verification & Validation of AI-Based Autonomous Systems
Abstract
As the deployment of AI agents in Automated Driving Systems (ADS) becomes increasingly prevalent, ensuring their safety and reliability is of paramount importance. In this talk, we present a novel approach to enhance the safety assurance of automated driving systems by employing formal contracts to specify and refine system-level properties. The proposed framework leverages formal methods to specify contracts that capture the expected behavior of perception and control components in ADS. These contracts serve as a basis for systematically validating system behavior against safety requirements during design and testing phases. We showcase the effectiveness of our approach through experiments conducted in the CARLA simulator, utilizing an off-the-shelf AI agent. By monitoring the components contracts on the ADS simulation, we could identify not only the cause of system failures, but also the situations which could lead to a system failure, facilitating the debugging and maintenance of the AI agents.
Presenter
Srajan Goyal
Poster 19
Authors and affiliations
Matteo Zavatteri University of Padova, Italy
Davide Bresolin University of Padova, Italy
Nicolò Navarin University of Padova, Italy
Title
Automated Synthesis of Certified Neural Networks
Abstract
Neural networks find applications in many safety-critical systems that raise concerns about their deployment: Are we sure the network will never advise doing anything violating a set of safety constraints? Formal verification has been recently applied to prove whether an existing neural network is certified for some property (i.e., if it satisfies the property for all possible inputs) or not. Formal verification can prove that a network respects the property, but cannot fix a network that does not respect it. In this work we focus on the automated synthesis of certified neural networks, that is, on how to automatically synthesize a network that is guaranteed to satisfy some required properties. We exploit a Counter Example Guided Inductive Synthesis (CEGIS) loop that alternates Deep Learning, Formal Verification, and a novel data generation technique that augments the training data to synthesize certified networks in a fully automated way. An application of a proof-of-concept implementation of the framework shows the feasibility of the approach.
Presenter
Davide Bresolin
Poster 20
Authors and affiliations
Federico Corò Department of Mathematics “Tullio Levi-Civita”, Università degli Studi di Padova, Italy
Lorenzo Palazzetti University of Perugia, Italy
Title
Pest Countermeasures Development Through Digital Twin Orchard Simulationsitle
Abstract
Protecting crops from the proliferation of emerging harmful insects is a critical challenge for modern agriculture. The use of Digital Twins (DTs) in agriculture is gaining momentum as an effective solution for assessing the impact of treatments against newly arising pests. In Italian orchards, the invasive species Halyomorpha halys (HH), commonly known as the brown marmorated stink bug, is threatening the entire pear production chain. Currently, there are no targeted pesticides available to combat this pest. As a result, broad-spectrum pesticides are widely used, which negatively impact the surrounding ecosystem, including pollinating insects. The Haly.ID project (www.haly-id.eu), launched in 2021, introduced initial tools and techniques for autonomously monitoring HH in the field using drones and IoT technology. Building on this foundation, this work aims to develop a DT capable of simulating complex interactions between HH and novel pesticides that are challenging to test in natural environments. The proposed DT will represent the monitored field using drone-captured data, incorporating digitalized instances of HH, fruit, and other key components essential to the pest’s lifecycle. Using insights from existing literature, a model of HH behavior will be integrated to simulate their movements over time and their responses to potentially new chemical agents. This project aspires to serve as a pilot prototype for extending the DT framework to include additional species and behavioral models. It offers a platform for developing countermeasures while simultaneously simulating their environmental impact, particularly on beneficial insects such as pollinators.
Presenter
Lorenzo Palazzetti
Poster 21
Authors and affiliations
Francesca Randone University of Trieste, Italy
Romina Doz University of Trieste, Italy
Cairoli Francesca University of Trieste, Italy
Luca Bortolussi University of Trieste, Italy
Title
Probabilistic Program for Collective Adaptive Systems
Abstract
The probabilistic programming paradigm is gaining popularity due to the possibility of easily representing probabilistic systems and running a number of off-the-shelf inference algorithms on them. This paper explores how this paradigm can be used to analyse collective systems, in the form of Markov Population Processes (MPPs). MPPs have been extensively used to represent systems of interacting agents, but their analysis is challenging due to the high computational cost required to perform exact simulations of the systems. We represent MPPs as runs of the approximate variant of the Stochastic Simulation Algorithm (SSA), known as 𝜏-leaping, which can be seen as a probabilistic program. We apply Gaussian Semantics, a recently proposed inference method for probabilistic programs, to analyse it. We show that 𝜏-leaping runs can be effectively analysed using a tailored version of Second Order Gaussian Approximation in which we use a Gaussian Mixture encoding of Poisson distributions. In the resulting analysis, the state of the system is approximated by a multivariate Gaussian Mixture generalizing other common Gaussian approximations such as the Linear Noise Approximation and the Langevin Method. Preliminary numerical experiments show that this approach is able to analyse MPPs with reasonable accuracy on the significant statistics while avoiding expensive numerical simulations
Presenter
Cairoli Francesca
Poster 22
Authors and affiliations
Giorgia Nadizar University of Trieste, Italy
Eric Medvet University of Trieste, Italy
Wilson G. Dennis University of Toulouse, France
Title
Interpretable Control Policies via Genetic Programming
Abstract
In most high-risk applications, interpretability is crucial for ensuring system safety and trust. However, existing research often relies on hard-to-understand, highly parameterized models, such as neural networks. In this paper, we focus on the problem of policy search in continuous observations and actions spaces. We leverage two graph-based Genetic Programming (GP) techniques—Cartesian Genetic Programming (CGP) and Linear Genetic Programming (LGP)—to develop effective yet interpretable control policies. Our experimental evaluation on eight continuous robotic control benchmarks shows competitive results compared to state-of-the-art Reinforcement Learning (RL) algorithms. Moreover, we find that graph-based GP tends towards small, interpretable graphs even when competitive with RL. By examining these graphs, we are able to explain the discovered policies, paving the way for trustworthy AI in the domain of continuous control.
Presenter
Eric Medvet
Poster 23
Authors and affiliations
Tagus Enes Atakli Department of Information Engineering and Mathematics, University of Siena, Italy
Pasquale Fedele Department of Information Engineering and Mathematics, University of Siena, Italy
Giovanna Maria Dimitri Department of Information Engineering and Mathematics, University of Siena, Italy
Title
Modelling a User-Centric Interactive AI Agent in Healthcare Applications: Towards Digital twin and Virtual counterparts
Abstract
This study develops a user-centric interactive AI agent utilizing advanced Natural Language Processing (NLP) techniques and GPT models to enhance personalized healthcare applications. By leveraging demographics and historical interactions, the agent significantly improves context-awareness and personalizes AI-generated responses. This approach addresses the prevalent under-utilization of both the capabilities of large language models and available user data. The system integrates a Natural Language Understanding (NLU) module, a Memory-Structure User Profile, a Decision Module, and a Prompt Adjusting Mechanism, demonstrating a critical use case in conversational AI for healthcare. The experimental setup included synthetic data generation, prompt comparison with the primary evaluation, and fine-tuning of the language models with the subsequent agent simulations. The research results indicated that enriched prompts incorporating contextual and user-specific information significantly boost the accuracy (up to 67.5 %), coherence (up to 72.27 %), personalization (up to 74.17 %), and relevance (up to 68.33 %) of responses across three distinct samples compared to prompts without additional information, with all improvements achieving statistical significance (p < 0.05). These advancements align the agent with the digital twin paradigm by serving as a virtual counterpart that maintains a dynamic user profile through real-time data integration and manages its lifecycle through integrated modules for creation, operation, and update to improve personalized and context-aware healthcare interactions.
Presenter
Giovanna Maria Dimitri
Poster 24
Authors and affiliations
Stefano Campanella OGS, Trieste, Italy
Luca Bortolussi University of Trieste, Italy
Stefano Querin OGS, Trieste, Italy
Stefano Salon OGS, Trieste, Italy
Title
Towards a data-driven ocean circulation model
Abstract
The accuracy of numerical weather prediction models for medium-term forecasts has steadily improved over the years. However, many human activities, such as extreme weather risk mitigation, require longer-term forecasts, which are inherently probabilistic due to the chaotic nature of the atmosphere. Subseasonal to seasonal forecasts address this need, traditionally relying on ensembles of physics-based atmosphere-ocean coupled simulations. More recently, data-driven models emerged as a competitive alternative. Still, these models lack a description of the ocean, whose coupling becomes critical at these timescales. We propose a data-driven model for skillful global ocean forecasting targeted at subseasonal-to-seasonal timescales and based on graph neural networks. We describe data preparation, motivate architectural choices and improvements based on analogies with finite element methods, and discuss training and validation procedures, with a focus on compute resource requirements and scalability.
Presenter
Stefano Campanella
Poster 25
Authors and affiliations
Dmitrii Kirov Collins Aerospace
Arthur Clavière Collins Aerospace
Title
Assurance of AI-Enabled Systems
Abstract
The use of AI components in safety-critical aviation systems (e.g., avionics) brings various benefits in terms of performance and enables new functionalities that are not possible to implement in traditional software. Such systems often require certification from global aviation authorities and, therefore, must exhibit high level of trust and guarantees on the absence of unintended behavior. This is specifically challenging for Machine Learning (ML) based systems (including Deep Learning and Reinforcement Learning), since ML models, such as neural networks, are complex and black box. Trustworthiness of AI/ML is achieved by providing design assurance, i.e., evidence that certain guidelines and verification processes have been followed during design and deployment.
This poster will present some of the research activities that Collins Aerospace is performing on the AI assurance topic. They include the use of Formal Methods for verification of safety-critical neural networks and their properties, such as stability, robustness, and generalization. Furthermore, scenario-based approaches for test generation will be presented. The poster will also discuss several AI-enabled aviation use cases of different complexity.
Presenter
Dmitrii Kirov, Ph.D. – Principal Research Engineer at Collins Aerospace
Poster 26
Authors and affiliations
Filippo Dalla Department of Physics and Astronomy "Augusto Righi", University of Bologna, Italy
Gregorio Berselli Department of Physics and Astronomy "Augusto Righi", University of Bologna, Italy
Dragos Dimitru Ioan Department of Physics and Astronomy "Augusto Righi", University of Bologna, Italy
Armando Bazzani Department of Physics and Astronomy "Augusto Righi", University of Bologna, Italy
Title
Towards an Urban Digital Twin: A use case for the optimization of urban traffic
Abstract
The concept of Digital Twin has been recently explored in the context of urban planning, in order to address problems with a degree of complexity otherwise difficult to manage. As a result, a noticeable amount of cities around the world are trying to bring this concept to reality, i.e. developing a Digital Twin of their entire urban area. Given that reproducing in detail the entities and the connections of an entire city is rather demanding, the goal of our research is to find the optimal level of detail (LOD) to build the Digital Twin of a city. Following some insights and results from complex systems physics, we argue that the effectiveness of such a project is not a linear function of the LOD, indeed having too many details (or parameters) can lead to overfitting, high parameter sensitivity and no added value. The LOD of a simulation is rather something to determine depending on the specific problems one want to address, especially when dealing with a Digital Twin, and retroaction enters the picture.
Starting from the development of a mesoscopic traffic simulation (a queuing model), we explicitly consider application to the Bologna road network using the traffic flow data available from magnetic coils to infer the traffic load, and a GPS data set on mobile phone positions that allows to reconstruct individual mobility paths and get information on the road network weights. We then built a Digital Twin use case by crafting a dynamic algorithm to optimize traffic light phases using real time data.
Presenter
Filippo Dalla
Poster 27
Authors and affiliations
Salvatore Cuomo University of Naples "Federico II", Italy
Federico Gatta Scuola Normale Superiore, Pisa, Italy
Vincenzo Vocca University of Naples "Federico II", Italy
Title
A Neural Network Approach for Liquidity Optimization in Automated Market Makers
Abstract
This study proposes a novel approach to liquidity provision in Decentralized Exchanges (DEXs), which leverage blockchain technology to facilitate asset exchanges without intermediaries. In contrast to Centralized Exchanges (CEXs), DEXs operate through Automated Market Makers (AMMs) and liquidity pools, thereby enabling users to provide or take liquidity. The present study focuses on Uniswap's v3 feature, Concentrated Liquidity (CL), in which liquidity providers (LPs) select specific price ranges for liquidity, thereby balancing risks and rewards. The present study combines parametric models to describe the dynamics of the market's principal components and the flexibility of neural networks. The value of the LP's position is regarded as a random variable, the expectation of which is obtained using the Feynman-Kac theorem and the Physics Informed Neural Network (PINN) to solve the corresponding Partial Differential Equation (PDE).The variance of future values can be obtained in a similar manner. Thus, it is possible to select the position range that optimizes the mean-variance utility function, as suggested in. As the primary computational burden of the neural network is in the offline training stage, our methodology is advantageous in terms of computational efficiency in the online evaluation stage, where multiple evaluations must be conducted within a limited time interval.
Presenter
Vincenzo Vocca
Poster 28
Authors and affiliations
Erik Chinellato University of Padova, Italy
Fabio Marcuzzi University of Padova, Italy
Title
Deep Unfolding for Embedded Digital Twins
Abstract
Deep neural networks (DNNs) have revolutionized numerous fields due to their powerful ability to learn complex representations. However, their black-box nature and lack of interpretability in architecture and weight design remain significant challenges. This poster presentation introduces the Deep Unfolding method as a promising alternative, bridging the gap between data-driven learning and model-based optimization. By unrolling iterative optimization algorithms into structured neural network architectures, Deep Unfolding provides a principled approach to network design, enabling interpretability and theoretical insights into their operation. We will explore how this method leverages domain knowledge, achieves faster convergence, and enhances performance in resource-constrained scenarios such as Embedded Digital Twins (EDTs). We will highlight a few wide-ranging practical applications of Deep Unfolding for EDTs, covering, e.g., audio source recognition and state estimation.
Presenter
Erik Chinellato
Poster 29
Authors and affiliations
Laura Meneghetti SISSA, Trieste, Italy
Edoardo Bianchi SISSA, Trieste, Italy
Nicola Demo SISSA, Trieste, Italy
Gianluigi Rozza SISSA, Trieste, Italy
Title
A Reduced Order Approach for ANNs applied to Image Recognition
Abstract
In the field of Deep Learning, the high number of parameters in models has become a significant concern within the scientific community due to the increased computational resources and memory required for training and inference. This naturally opens several computational issues when these deep learning models are deployed in vision embedded systems. This kind of vision devices are often endowed with limited hardware resources, strict memory constraints, and with a low CPU performance. Addressing this issue, we propose a novel compression methodology, employing Model Order Reduction techniques or Tensor Decompositions to perform reduction. In this way, we aim to develop an approach based on these mathematical tools, detecting the most relevant parameters of a model in order to construct a compressed version of it. Specifically, we replace certain layers of the original architecture with layers that perform linear projections onto a reduced space defined by our reduction technique. The compressed network is then trained through a novel Knowledge Distillation approach. We conducted experiments on image classification tasks using multiple architectures and datasets. The results underscore the effectiveness of our method in significantly reducing the number of parameters and the overall size of neural networks while maintaining high performance.
Presenter
Laura Meneghetti
Poster 30
Authors and affiliations
Kabir Bakhshaei SISSA, Trieste, Italy
Sajad Salavatidezfouli SISSA, Trieste, Italy
Giovanni Stabile SISSA, Trieste, Italy
Gianluigi Rozza SISSA, Trieste, Italy
Title
Boundary Velocity Profile Prediction for Patient-Specific Cardiovascular Flow Modeling via Stochastic Data Assimilation
Abstract
Accurate velocity boundary profiles are critical for high-fidelity simulations of cardiovascular flows, as they directly impact wall shear stress predictions—key in diagnosing diseases like atherosclerosis. However, in-vivo data from modalities like 4D flow MRI often suffer from noise and low resolution. We address this challenge using a stochastic data assimilation approach that integrates computational fluid dynamics with an Ensemble-based Kalman filter to iteratively refine unknown boundary conditions in real-time. Applying the incompressible Navier-Stokes equation, we tested constant, time-dependent, and space-time-dependent boundaries in 2D and 3D vascular models. This work enhances the reliability of cardiovascular simulations for better diagnosis and treatment of diseases.
Presenter
Kabir Bakhshaei
Poster 31
Authors and affiliations
Aldo Canfora Department of Physics and Astronomy, University of Bologna, Italy
Armando Bazzani Department of Physics and Astronomy, University of Bologna, Italy
Mirko Degli Esposti Department of Physics and Astronomy, University of Bologna, Italy
Title
Urban Building Energy Modelling
Abstract
For European Commission, the Urban Digital Twin is crucial to addressing challenges such as decarbonization and achieving the Green Deal goals and to develop Smart Cities. Infrastructures energy management makes possible to reduce consumption, integrate renewable energy and plan sustainable cities. This technology is a strategic lever for creating greener, more resilient cities aligned with European sustainability goals.
Urban building energy modeling (UBEM) is a key tool for analyzing and optimizing energy consumption at the urban level. Using geometric data from LiDAR surveys and open data, three-dimensional models of buildings can be generated. These models are integrated with the EnergyPlus simulation engine to analyze the energy consumption and thermal behavior of buildings, considering morphological parameter and climate variables. This approach enables the evaluation of energy efficiency and sustainability strategies to support urban planning and the decarbonization of cities.
Presenter
Aldo Canfora
Poster 32
Authors and affiliations
Lorenzo Bonin University of Trieste, Italy
Federica Adobbati OGS, Trieste, Italy
Stefano Campanella OGS, Trieste, Italy and University of Trieste, Italy
Andrea De Lorenzo University of Trieste, Italy
Luca Manzoni OGS, Trieste, Italy and University of Trieste, Italy
Title
Super-resolution of Global Ocean Physics Reanalysis data with Diffusion Models
Abstract
Ocean physics reanalysis data provide critical insights into the state of the ocean by integrating observations and numerical models. Moreover, uncertainty quantification and robust predictions remain significant challenges due to the complexity and variability of oceanic systems. Traditional ensemble numerical models play a crucial role in addressing these challenges by capturing a range of possible outcomes, thus providing a more comprehensive representation of uncertainties. On the other hand, their computational demands impose severe constraints on resolution, limiting their applicability. To address these challenges, deep learning approaches are increasingly gaining popularity for super-resolution tasks in geoscience. Among these, diffusion probabilistic models have recently demonstrated state-of-the-art performance in many different domains, albeit with significant memory limitations at large scales. In this work, we propose a deep learning approach for enhancing the resolution of global ocean physics reanalysis data with memory-efficient diffusion models. The model is trained and validated on historical data, showing that it can recover high-resolution features typically missed by traditional interpolation models.
Presenter
Lorenzo Bonin