Isabel Valera

I wrote a new paper that I suspect might become significant: https://lnkd.in/epB3HqFk Concept-Based Generic Programming in C++ Abstract We present programming techniques to illustrate the facilities and principles of C++ generic programming using concepts. Concepts are C++’s way to express constraints on generic code. As an initial example, we provide a simple type system that eliminates narrowing conversions and provides range checking without unnecessary notational or run-time overheads. Concepts are used throughout to provide user-defined extensions to the type system. The aim is to show their utility and the fundamental ideas behind them, rather than to provide a detailed or complete explanation of C++’s language support for generic programming or the extensive support provided by the standard library. Generic programming is an integral part of C++, rather than an isolated sub-language. In particular, key facilities support general programming as well as generic programming (e.g., uniform notation for types, lambdas, variadic templates, and C++26 static reflection). Finally, we give design rationales and origins for key parts of the concept design, including use patterns, the relationship to Object-Oriented Programming, value arguments, notation, concept type-matching, and definition checking.

2.408

84 Kommentare

🚀 New Preprint Alert! 📄 We are thrilled to announce the release of our latest preprint: “Port-Hamiltonian Neural Networks: From Theory to Simulation of Interconnected Stochastic Systems” by Luca Di Persio, Matthias Ehrhardt, Youness Outaleb, and Sofia Rizzotto 🎓 🔗 Read the preprint here: https://lnkd.in/dWceDxNz 🔍 What’s inside? This work introduces Port-Hamiltonian Neural Networks (pHNNs) 🧠⚡, a novel framework combining: 🏗️ Port-Hamiltonian Systems (PHS) for energy-consistent modeling 🤖 Neural Networks for data-driven learning 🌪️ Stochastic dynamics to handle uncertainty and noise Our approach bridges the gap between deterministic and stochastic modeling, ensuring stability through passivity-preserving structures and enabling accurate simulations of interconnected stochastic systems. 🧪 Applications & Highlights ✅ Advanced modeling of nonlinear and stochastic systems ✅ Simulation results on damped mass-spring systems, chaotic Duffing oscillators, and robotic control ✅ Robust learning of complex dynamics, even under high noise ✅ A solid foundation for data-driven control of multi-agent and robotic systems 🤝 Collaboration & Context This work is part of an exciting collaboration between the University of Verona 🇮🇹 and the University of Wuppertal 🇩🇪, carried out within Project C05 of the newly established DFG Collaborative Research Centre (CRC 1701) on Port-Hamiltonian Systems 🎯, see https://lnkd.in/dK5K7CtV 📌 Citation Di Persio, L., Ehrhardt, M., Outaleb, Y., & Rizzotto, S. Port-Hamiltonian Neural Networks: From Theory to Simulation of Interconnected Stochastic Systems. IMACM Preprint 25/15, September 2025. 🌟 We are excited to share this milestone and look forward to your feedback! #Preprint #Research #NeuralNetworks #PortHamiltonian #StochasticSystems #MachineLearning #ControlTheory #DFG #CRC1701 #Wuppertal #Verona

93

3 Kommentare

Happy to share that our new paper "Uncertainty-Aware and Reliable Neural MIMO Receivers via Modular Bayesian Deep Learning", was accepted for publication in the IEEE Transactions on Vehicular Technology. https://lnkd.in/dmZiJX_3 The work, lead by Tomer Raviv in collaboration with Sangwoo Park and Osvaldo Simeone, shows the usefulness of using Bayesian deep learning combined with modular model-based deep learning architectures in the context of deep receivers. While Bayesian deep learning is typically about improving calibration, here improving calibration actually improves performance (due to propagation of more faithful soft estimates). And what more can we get by combining Bayesian learning with model-based deep learning? - rapid online learning (by casting learning as Bayesian tracking). - unsupervised drift detection (leveraging Bayesian uncertainty measures). But these are topics for future works (and future posts....)

101

At Delft Institute of Applied Mathematics, recently an open textbook on Linear Algebra was published within the context of PRIME (Program Innovation in Mathematics Education). The principal authors are Nikolaas Verhulst and Andre Hensbergen. It is an introduction to one of the founding areas of machine learning. If you already know the content, still it is great to see how the writing team has chosen to bring the subjects supported by visualizing applets:  https://lnkd.in/eUU_qKnW TU Delft Interactive Textbooks on a variety of subjects can be found at https://lnkd.in/emFVy2Uw #tudelft #mathematics #machinelearning #openeducation

200

8 Kommentare

New (another) Nature Communications paper (https://lnkd.in/g2qmRAuy) We continue our journey on optical computing. In this paper, we present an optical neural engine (ONE) architecture combining diffractive optical neural networks (DONNs) for Fourier space processing and optical matrix multipliers for real space processing to solve time-dependent and time-independent partial differential equations (PDEs) in diverse disciplines, including Darcy flow equation, the Poisson’s equation in demagnetization, the Navier-Stokes equation in incompressible fluid, Maxwell’s equations in nanophotonic metasurfaces, and coupled PDEs in a multiphysics system. We numerically and experimentally demonstrate the capability of the ONE architecture. This architecture not only leverages the advantages of dual-space processing but also can be implemented using optical computing hardware with low-energy and highly parallel constant-time processing. The demonstrated architecture offers a versatile and powerful platform for large-scale scientific and engineering computations. Thank all authors for their great efforts. Thank NSF for the support.

111

6 Kommentare

Design and Automation of Shuttling in Trapped-Ion Quantum Computing! This is the topic of Daniel Schoenberger's dissertation project — and the focus of his presentation this week at the Student Research Competition #SRC of ICCAD.🎓 ⚛️ In #TrappedIon #QuantumComputers, #shuttling — the movement of ions between memory and processing zones — is essential for executing #QuantumCircuits. But it's also a complex and error-prone task that must be optimized to preserve quantum coherence and reduce execution time. Daniel's work tackles this challenge head-on by proposing automatic #compilation and #scheduling methods that orchestrate ion movement across multi-zone #QCCD architectures — enabling scalable and efficient #QuantumComputing. His contributions include: 📄 #Publications at top venues: #TCAD, Design, Automation and Test in Europe Conference (DATE), #ASPDAC , #IEEEQuantumWeek, and #QSW 🧰 An #OpenSource implementation: The #MQT #IonShuttler (check out at https://lnkd.in/eRp8jmmk) 🖥️ A #visualization tool (as seen on the tablet in the photo!) that helps researchers understand and optimize ion movement 👏 Huge congratulations to Daniel for pushing the boundaries of #DesignAutomation and #QuantumComputing — and for making these innovations accessible to the community through open #tools and reproducible research. #ICCAD2025 #EDA #CAD #QuantumComputing #TrappedIon #QCCD #IonShuttling #MQT #OpenSource

102

1 Kommentar

#new #quantum #SciML #work out https://lnkd.in/eRBVx-SK — and we are now closer to the goal of building quantum pipelines for studying engineering problems. In the recent study we have tried to answer the question: How can we build algorithms that can extract patterns and relevant features from quantum data? Flow fields in quantum #CFD simulations are prepared as states in a Hilbert space — something you don't just plot with #matplotlib... Specifically, how do you know if a "vortex was there?" Turns our these are properties one can extract from quantum Fourier spectra in a special way, being simultaneously meaningful and efficiently measuable. While still being a baby step in the pipeline, we see that quantum solutions to different equations reveal properties that we care, with a benefit of coherent processing, as long as we ask nicely. Results are summarised in https://lnkd.in/eRBVx-SK , check it out. Huge thanks to Chelsea Williams who took on the challenge and didn't give up while trying various approaches we brainstormed. And of course our collaboration with Siemens Digital Industries Software with Daniel Berger leading the way, and amazing collaborators Annie Paine Antonio Andrea Gentile

118

1 Kommentar

Machine Learning for Materials Informatics - Jul 28 - 31, 2025 (3.5 days) - LIVE VIRTUAL Instructor: Prof. Markus J Buehler, mbuehler@MIT.EDU    Material informatics is transforming the way materials are discovered, understood, developed, selected, and used. In this condensed course, you will engage in interactive lectures, clinics, and labs designed to help you learn, design, and apply modern material informatics tools and large-scale multiscale modeling - with the ultimate goal of helping you to speed up your design process and implement cost effective rapid discovery and prototyping in your organization. With data available from autonomous experimentation or large databases and data mining, there exist many opportunities to accelerate and expand your materials design platform. The course involves a mix of lectures, hands-on labs and clinics for an immersive experience. Participants will learn fundamentals and techniques to deploy machine learning in materials development and gain first-hand understanding of state-of-the art tools for varied applications ranging from data mining to inverse design. Topics of emergent interest include multimodal LLMs, multi-agent modeling, and self-driving discovery and other exciting new developments in AI for science and discovery. Learn more here: https://lnkd.in/gsPsjy_j (link to sign-up)

165

9 Kommentare

We are happy to welcome Maike Meier to the Bernoulli Institute, in the Computational and Numerical Mathematics group! Maike's research interests are broadly numerical linear algebra, which is the study of efficient and accurate computer algorithms for matrix problems. She interested in using approximate computing techniques, such as randomisation and mixed-precisions, for linear problems. These approaches can be employed to improve the computational feasibility of solvers for very large-scale problems. More recently, she has been working with scientists to use ideas from numerical linear algebra to accelerate imaging experiments. Maike believes there's a wealth of potential for interdisciplinary research which combines experimental techniques with modern NLA concepts, beyond reducing computational effort.

160

6 Kommentare

Hi all, I’m trying something new — I’ve just published my first research blog post as a way to share and discuss some of my work. The post dives into our approach for inferring underlying pore-scale dynamics from pressure measurements during core-scale experiments. Traditionally, directly imaging residually trapped CO₂ in the pore space is time-consuming, expensive, and limited by sample size. But by analysing the spectral power of pressure signals using continuous wavelet transformation (CWT), we can capture rich information about the evolution of flow — without relying solely on imaging. The key takeaway: averaging pressure data over time can wash away important details about the fluid dynamics happening inside the core. Our method helps reveal those hidden processes. If you’re curious about how pressure data can tell a deeper story about CO₂ trapping, you can read the blog post here: https://lnkd.in/giBnMdc5 The code is available here: https://lnkd.in/gp9_nyzW A huge thanks to Katriona Edlmann and Gareth Roberts for their help in this work! #Geoscience #CarbonCapture #CCUS #PorousMedia #WaveletAnalysis #ScienceCommunication

125

9 Kommentare

Happy to see our paper “Factor Models with Sparse Vector Autoregressive Idiosyncratic Components” out at Oxford Bulletin of Economics and Statistics! This is a joint work with the great Jonas Krampe, to whom I am very grateful for these many years of great work together! What do we do? We challenge the widespread misconception that factor models are per-se a dimensionality reduction technique (PCA is!) and instead we advocate the explicit modeling of the high-dimensional vector of idiosyncratics via regularization. In fact, the factor model representation, “common + idiosyncratic components”, of a high-dimensional vector of time series, was envisioned to “divide and conquer”. The low-rank common component, driven linearly by a finite, small number of factors, represents the co-movements among the series; instead, the idiosyncratic component contains the individual features of the series and possibly measurement error. These two components are different objects, which should be modeled differently to better capture their own dependencies. If the idiosyncratic component would truly contain only measurement error, its modeling might be unnecessary (this might happen e.g., for some macroeconomic series). However, If it does not, then its modeling is paramount to avoid leaving some cross-sectional and time dependence unattended. Regularized estimation of the high-dimensional vector of idiosyncratics is sensible, since once the factors have been accounted for, it is much more likely for the left-overs to then be sparse. We indeed employ a high-dimensional, approximate static factor model and assume the idiosyncratic term follows a sparse vector autoregressive model (VAR). To estimate this we require two steps: (i) factors and loadings are estimated via principal component analysis (PCA); (ii) a sparse VAR is estimated via (adaptive) lasso on the estimated idiosyncratic components from (i). Step (ii) is what allows to model cross-sectional and time dependence left after the factors estimation. We prove the consistency of this approach as the time and cross-sectional dimensions diverge. In (ii), we allow sparsity to be very general: approximate, row-wise, and growing with the sample size. However, step (ii) is performed on the estimatedidiosyncratics coming from (i). Hence, one inescapably has some error carrying forward from (i) to (ii) to account for. We derive a refined expression of such error, which enables us to obtain tight rates for the lasso in (ii). We discuss applications on forecasting & factor-augmented regression and present an empirical application on macroeconomic forecasting using the Federal Reserve Economic Data - Monthly Database (FRED-MD). The latter application suggests how for some important series such as inflation and industrial production the proposed approach can yield better forecasting performances compared to traditional methods. Check it out! https://lnkd.in/djpHJiJE

96

1 Kommentar