Publications

itwinai: A Python Toolkit for Scalable Scientific Machine Learning on HPC Systems

Published in Journal of Open Source Software, 2026

The integration of Artificial Intelligence (AI) into scientific research has expanded significantly over the past decade, driven by the availability of large-scale datasets and Graphics Processing Units (GPUs), in particular at High Performance Computing (HPC) sites. However, many researchers face significant barriers when deploying AI workflows on HPC systems, as their heterogeneous nature forces scientists to focus on low-level implementation details rather than on their core research. At the same time, the researchers often lack specialized HPC/AI knowledge to implement their workflows efficiently.

Recommended citation: Bunino et al., (2026). itwinai: A Python Toolkit for Scalable Scientific Machine Learning on HPC Systems. Journal of Open Source Software, 11(117), 9409, https://doi.org/10.21105/joss.09409 https://doi.org/10.21105/joss.09409

itwinai: Enabling Scalable AI Workflows on HPC for Digital Twins in Science

Published in 27th International Conference on Computing in High Energy and Nuclear Physics (CHEP 2024), 2025

itwinai is a Python library designed to facilitate scalable AI workflows on High-Performance Computing (HPC) systems. By abstracting complex engineering tasks, it reduces overhead and enables seamless scaling of AI models across diverse infrastructures. Deployed on HPC systems, such as Jülich and Vega, itwinai allows users to switch between distributed machine learning (ML) frameworks, including PyTorch DDP, DeepSpeed, and Horovod, through a unified interface. The library integrates mechanisms for profiling computational efficiency, tracking key metrics, such as GPU utilization and power consumption, and optimizing hyperparameters at scale. Additionally, it provides transparent offloading of compute-intensive tasks from the cloud to HPC systems and supports model and pipeline parallelism to accommodate largescale models. A continuous integration and deployment pipeline ensures reproducibility and compatibility across environments. In the interTwin project, itwinai has been utilized in physics and climate research use cases, demonstrating its potential to enhance sustainability and computational efficiency. Its integration with leading scientific computing centers highlights its role in advancing AI-driven digital twins and addressing large-scale scientific challenges.

Recommended citation: Bunino, Matteo, Jarl Sondre Sæther, Anna Elisa Lappe, Maria Girone, and Kalliopi Tsolaki. “Itwinai: Enabling Scalable AI Workflows on HPC for Digital Twins in Science.” EPJ Web of Conferences 337 (2025): 01361. https://doi.org/10.1051/epjconf/202533701361. https://doi.org/10.1051/epjconf/202533701361

A software ecosystem for multi-level provenance management in large-scale scientific workflows for AI applications

Published in SC24-W: Workshops of the International Conference for High Performance Computing, Networking, Storage and Analysis, 2024

Scientific workflows and provenance are two faces of the same medal. While the former addresses the coordinated execution of multiple tasks over a set of computational resources, the latter relates to the historical record of data from its original sources. This paper highlights the importance of tracking multi-level provenance metadata in complex, AI-based scientific workflows as a way to (i) foster and (ii) expand documentation of experiments, (iii) enable reproducibility, (iv) address interpretability of the results, (v) facilitate performance bottlenecks diagnosis, and (vi) advance provenance exploration and analysis opportunities.

Recommended citation: G. Padovani et al., "A software ecosystem for multi-level provenance management in large-scale scientific workflows for AI applications," SC24-W: Workshops of the International Conference for High Performance Computing, Networking, Storage and Analysis, Atlanta, GA, USA, 2024, pp. 2024-2031, doi: 10.1109/SCW63240.2024.00253 https://doi.org/10.1109/SCW63240.2024.00253

From Spectral Graph Convolutions to Large Scale Graph Convolutional Networks

Published in arXiv preprint arXiv:2207.05669, 2022

Graph Convolutional Networks (GCNs) have been shown to be a powerful concept that has been successfully applied to a large variety of tasks across many domains over the past years. In this work we study the theory that paved the way to the definition of GCN, including related parts of classical graph theory. We also discuss and experimentally demonstrate key properties and limitations of GCNs such as those caused by the statistical dependency of samples, introduced by the edges of the graph, which causes the estimates of the full gradient to be biased. Another limitation we discuss is the negative impact of minibatch sampling on the model performance. As a consequence, during parameter update, gradients are computed on the whole dataset, undermining scalability to large graphs. To account for this, we research alternative methods which allow to safely learn good parameters while sampling only a subset of data per iteration. We reproduce the results reported in the work of Kipf et al. and propose an implementation inspired to SIGN, which is a sampling-free minibatch method. Eventually we compare the two implementations on a benchmark dataset, proving that they are comparable in terms of prediction accuracy for the task of semi-supervised node classification.

Recommended citation: Matteo Bunino (2022). "From Spectral Graph Convolutions to Large Scale Graph Convolutional Networks" Journal arXiv preprint arXiv:2207.05669. https://arxiv.org/pdf/2207.05669

Reinforcement Learning-aided Dynamic Analysis of Evasive Malware

Published in PoliTO, 2022

As cyberspace becomes more and more complex, malware authors strive to take advantage of the growing number of vulnerabilities. This requires security researchers to invest more effort in developing automated malware analysis tools, able to cope with the increasing pace of suspicious binaries. The Achilles heel of automated malware analysis tools is evasive malware, which may put countless strategies in place to impede analysis. For instance, malware trying to detect that they are under observation in an analysis environment and, as a consequence, conceal their malicious behavior by performing only innocuous operations. Evasive malware can hinder analysis by either performing static code obfuscation (eg via packers, crypters) or dynamic evasion (eg sandbox and debugger evasion). The goal of this thesis is to explore the application of Reinforcement Learning (RL) to dynamic analysis, to reduce the burden of exhaustive exploration of conditional paths. To this end, we develop a model that is capable of noticing new evasive schemes via RL. Stateof-the-art approaches generally identify evasions through fingerprinting, which is not effective in identifying slight mutations of evasion schemes. In contrast, this work employs a language model to abstract out the syntax of binary code, while preserving its semantics. Furthermore, to improve over state-of-the-art solutions, the presented solution considers both evasion schemes and the malicious nature of regions of code protected by evasive conditions, to better distinguish true evasive behaviors from false positives. As a consequence, this method is easily extended to guide the search of…

Recommended citation: Matteo Bunino (2022). "Reinforcement Learning-aided Dynamic Analysis of Evasive Malware. https://webthesis.biblio.polito.it/secure/22588/1/tesi.pdf