Computer Science Joint Seminar 2025

Welcome by Opening

Marco Aiello

Towards Graph Foundation Models for Water Distribution Networks

Andrés Tello

Abstract: We present the first approach towards Foundation Models for Water Distribution Networks (WDNs). Foundation models are a special type of deep learning models trained on diverse and vast amounts of data to solve generic tasks. Then, the models are adapted to tackle specific problems from a particular domain. We show that WDN research can also benefit from these technological advancements to boost scientific exploration and optimize day-to-day WDN operations. Current Deep Learning methods for Graphs, tailored to the water domain, can be the enablers of Graph Foundation Models for WDNs. These models are trained on generic tasks to learn more accurate data representations based on the topology of WDNs and the properties of junctions and pipes. In this work, we address the state estimation problem in WDNs to showcase the benefits of Graph Foundation Models with respect to models trained for a specific task on a single WDN.

Windowed Multivariate LSTM Autoencoder for Leak Detection in Water Distribution Networks

Samer Ahmed

Abstract: Leakages in water distribution networks (WDNs) are highly challenging to accurately identify and label, leading to significant water loss and high repair costs. This research proposes a Windowed LSTM-based multivariate Autoencoder for detecting leakages. For evaluation, we compare it to a 1D CNN Autoencoder trained in a similar manner. We evaluate the two models in terms of normal reconstruction error, factors of mean reconstruction error (a novel metric), and time to detection. In particular, we focus on detecting incipient leaks in DMA C of the BattleDIM dataset, which has irregular demand patterns. The Windowed LSTM Autoencoder had the lowest normal reconstruction error, the highest error deviations when leaks were present, and detected incipient leaks fastest.
We also discuss ongoing work regarding adding physics constraints to leak detection models.

GenAI-based test case generation and execution in SDV platform

Denesa Zyberaj

Abstract: A GenAI-driven approach for automated test case generation, leveraging Large Language Models and Vision-Language Models to translate natural language requirements and system
diagrams into structured Gherkin test cases. The methodology integrates Vehicle Signal Specification modeling to standardize vehicle signal definition, improve compatibility across automotive subsystems, and streamline integration with third-party testing tools. 

Adapt or Collapse: Graph Neural Networks at Test Time

Huy Truong

Abstract: Deploying Graph Neural Networks (GNNs) in machine-learning pipelines benefits graph-based systems like water distribution networks, but fixed model weights often degrade under distribution shifts. A potential approach is Test-Time Training (TTT), which enables model adaptation during inference by leveraging unlabeled test data with a Self-Supervised Learning (SSL) task. However, selecting a robust SSL task for GNNs remains a challenge, and TTT typically updates only limited parts of the model. To resolve this, we systematically identify an effective SSL task and propose T3R, which combines TTT with Rotograd, a method that aligns task gradients via rotation matrices. During inference, T3R utilizes these matrices and SSL signals to produce surrogate gradients, enabling broader model adaptation from test inputs. Experiments on out-of-distribution datasets show T3R outperforms standard inference. In addition, we observe that the TTT-based approach's performance positively correlates with gradient alignment between tasks, offering a useful indicator for auxiliary task selection.

Realistic AI Planning: Risk Awareness in HTNs, Applications, and Beyond

Ebaa Alnazer

Abstract: AI planning in real-world domains requires models that capture their realistic aspects, such as uncertainty and risk, yet these are often oversimplified. To address this, we introduce a conceptual framework that identifies and categorises planning aspects, providing guidance for designing AI planning systems. We illustrate this by systematically modelling autonomous vehicles and satellites to better capture real-world complexity and move domains closer to reality. Central to this is risk-awareness, where we enhance Hierarchical Task Network (HTN) planning with concepts from expected utility theory by modelling risk and uncertainty through probability distributions of action costs. This allows defining risk-aware HTN planning as capable of accounting for different risk attitudes rather than assuming risk neutrality, by computing plans with the highest expected utility. We also highlight the interplay between trust and risk in autonomous driving and propose addressing the trust-based driving task as an AI planning problem that aligns with risk-aware HTN planning to explicitly reflect driver trust. This work forms a stepping stone towards AI planning approaches applicable to real-world settings.

Toward Graph-Based Intelligence In Wastewater Systems

Revin Naufal Alief

Abstract: Graph Neural Networks (GNNs) offer a powerful framework for learning from network-structured data by capturing both local characteristics and system-wide interactions. This makes them highly suitable for modeling urban wastewater systems, where spatial and structural dependencies are essential. In this work, we tried a practical pipeline that integrates InfoWorks simulation data with GNNs to estimate missing water-level values across a sewer network. Our case study focuses on a medium-sized system with 894 nodes and 968 links, where nodes represent components such as manholes, storage areas, and outfalls, while links correspond to pipes, channels, or orifices. Missing sensor data is simulated using a fixed masking scenario to reflect challenges such as partial sensor coverage or failure. The GNN performance achieves mean absolute errors between 0.0157 m and 0.351 m, with higher errors near flow control structures like weirs and pumps. These results demonstrate the potential of GNNs not only for accurate estimation but also for uncovering spatial patterns of uncertainty, offering a foundation for further model improvement.

LLM for SOC

Robin Pesl

Abstract: TBD

On the feasibility of identifying microservice early-stage architectures using LLMs

Marco Calamo 

Abstract: Microservice architectures have gained prominence for their modularity, scalability, and flexibility, particularly when paired with Domain Driven Design to guide service boundaries and decomposition. Also, recently, AI-assisted tools based on Large Language Models (LLMs) are extensively used in the world of software development. However, such tools for software development focus mostly on low-level code tasks, offering limited support for high-level architecture design. This presentation will introduce ArchiLLM, a tool leveraging LLMs to assist software designers in the early stages of microservice architecture design. Given minimal initial input, ArchiLLM recommends appropriate microservices and relevant data-centric design patterns (e.g., CQRS, Saga, API Composition, and Event Sourcing). We will validate the tool using the newly developed Archi Dataset, which comprises academic microservice projects and a dataset of well-known open-source microservice projects. Both automated metrics and expert evaluations confirm ArchiLLM’s effectiveness in supporting architectural decision-making.

Machine Learning as a Service

Adriano Puglisi

Abstract: Machine Learning as a Service (MLaaS) has emerged as a paradigm that enables organizations to access advanced artificial intelligence capabilities without the need to build and maintain complex infrastructures. By exposing machine learning models and pipelines through cloud–based interfaces, MLaaS abstracts away the technical challenges of data preprocessing, model training, deployment, and scalability. This approach fosters rapid experimentation and lowers the entry barrier for sectors where computational resources and specialized expertise are limited. At the same time, the integration with edge devices extends the service to distributed environments, supporting real–time inference and adaptive retraining closer to data sources.

Microservices in Constrained Environments

Jacopo Rossi

Abstract: Microservice architecture is widely recognized as a paradigm that enhances scalability, maintainability, and the evolution of software systems. Adopting this approach in resource- and time-constrained environments introduces unique challenges that require careful consideration of trade-offs between architectural benefits and practical limitations. Large Language Models (LLMs) can provide valuable assistance in this process, supporting the definition of architectural requirements by extracting and synthesizing knowledge from technical materials and best practices. The complexity of coordinating interconnected microservices under strict temporal and resource constraints creates a valuable opportunity to develop automated workload scheduling and deployment technologies. This presentation provides an overview of the benefits and challenges of using microservices in constrained environments, highlighting future directions for research and practical implementation.

Opportunities for Improving the GPU Utilization of Deep Learning workloads: Insights from the PyTorch Backend.

Mahmoud Alasmar

Abstract: Modern GPU clusters are designed to meet the computational demands of Deep Learning (DL) workloads. However, recent studies reveal a persistent gap between the computational capacity offered by GPU clusters and the actual utilization of resources by DL workloads, leading to wasted computational power, operational costs, and energy. Existing techniques, such as GPU sharing and elastic scheduling, improve efficiency but fail to fully address bottlenecks that emerge at the granularity of GPU kernels. In this work, we analyze the execution flow of running a DL model from the backend perspective of PyTorch. Using the roofline analysis, we identify performance-limiting operations and opportunities for improving GPU utilization. Our findings reveal how PyTorch backend behavior influences GPU usage efficiency and highlight areas for further optimization.