Stefano Riva

Stefano Riva

Researcher in computational methods for nuclear reactors, with a focus on Reduced Order Modelling and Machine Learning for state estimation. Passionate about bridging physics-based models and data-driven approaches.

Stefano Riva’s Webpage

Current position: Researcher at the Department of Energy, Politecnico di Milano.

I am Stefano Riva, a researcher in computational methods for nuclear reactors. My work focuses on advanced numerical modelling, reduced order methods, and scientific machine learning to improve the efficiency, safety, and reliability of reactor simulations.

I am currently working at Politecnico di Milano, where I work on data-driven and physics-based approaches for reactor dynamics and state estimation. My research integrates high-performance computing, multi-physics modelling, and machine learning to address complex challenges in modern nuclear engineering.

I am the developer of several open-source frameworks, including pyforce, pySHRED, and ctf4science, focusing on model reduction and scientific machine learning for dynamical systems.

Research Interests

  • Scientific Machine Learning techniques for Nuclear Reactors
  • Nuclear Reactor Modelling
  • Reduced Order Modelling and Data Assimilation
  • Computational Fluid Dynamics
  • Numerical Methods for engineering

PhD Thesis

📘 PhD Repository: github.com/Steriva/PhD-Thesis

Title: Advanced Data-Driven Techniques for State Estimation in Nuclear Reactors

Supervisors: Prof. Antonio Cammi, Dr. Carolina Introini, Prof. J. Nathan Kutz

I got a PhD in Nuclear Engineering in 2025 at Politecnico di Milano (Italy), focused on the development of fast, reliable, and data-driven state estimation techniques for nuclear reactors, with a particular emphasis on next-generation concepts such as molten-salt systems. These reactors pose significant sensing challenges due to their harsh operating environments, making traditional in-core measurements difficult or impossible.

To address this, my research investigated how to infer the full reactor state from few, sparse, and often indirect sensor observations, combining them with physics-based background models described by parametric PDEs within a Data Assimilation framework combined with Reduced Order Modelling (ROM) to enable real-time computation. Classical ROM approaches—such as GEIM and the Parametrized-Background Data-Weak formulation—were explored, as well as advanced data-driven methods capable of correcting model inaccuracies. Building on recent developments in Scientific Machine Learning, I extensively studied the Shallow Recurrent Decoder (SHRED) architecture, a lightweight neural network framework capable of handling nonlinear dynamics, parametric datasets, and complex reactor behaviour with minimal training effort.

These techniques were validated on several challenging case studies, including:

  • Molten Salt Fast Reactor (MSFR)
  • TRIGA Mark II research reactor
  • Fluid-dynamics benchmarks
  • The DYNASTY experimental facility at Politecnico di Milano

The results demonstrate the strong potential of ROM-based and machine-learning-based state estimation methods for future digital twins of nuclear systems, ultimately supporting safer and more efficient reactor operation.