FEMTO-ST Institute

This study investigates the use of physics-augmented neural networks to predict the frequency-response behavior of a nonlinear piezoelectric vibration energy harvester (PVEH). An analytical model is first developed under the assumption of geometrical linearity, incorporating a Duffing-type nonlinearity arising from the magnetic interaction, referred to as the baseline model. The physical parameters are calibrated using Particle Swarm Optimization (PSO) and the PVEH’s behavior is experimentally validated, based on frequency response curves collected from low to high acceleration levels. While the baseline model predicts the PVEH behavior accurately at low accelerations, it fails at higher accelerations where nonlinear effects become significant. To extend the model’s range of applicability, a physics-augmented neural network is introduced, resulting in an augmented model that integrates physical knowledge through loss and activation functions. Unlike existing data-driven approaches that typically focus on isolated response features, the proposed framework enables prediction of the frequency response in the presence of nonlinearities not accounted for in the analytical development. Performance evaluation within the -3 dB bandwidth region shows that the mean absolute percentage error (MAPE) of the baseline model is 81.1

Guided elastic waves are a truly cross-disciplinary key enabling technology. For more than five decades, surface acoustic wave (SAW) and bulk acoustic wave devices find widespread applications. Nowadays, different types of guided elastic waves cover the wide spectrum of applications spanning from quantum technologies to the life sciences, from controlling single excitations to macroscopic collective states in condensed matter. Six years after the first 2019 SAW roadmap, we believe it is time to make a step back and take a fresh look at the status of the field and its future challenges. Since the first roadmap in 2019, the spectrum clearly expanded and this new edition presents a current snapshot of the status of this vibrant field and prospects for potential future developments.

Synthetic data generation has become a key solution for enabling data sharing, accounting for privacy constraints, and addressing the limited availability of real-world datasets, particularly in tabular format. Conventional statistical models and generative adversarial networks are widely used, but they require dataset-specific preprocessing and sometimes struggle to capture complex semantic relationships among features accurately. Recent advances in large and small language models(LLMs and SLMs) have improved the development of synthetic tabular data by introducing transformer-based architectures and textual representations of tables. This paper reviews and analyzes in detail language model-based methods for generating synthetic tabular data, focusing on state-of-the-art approaches such as GReaT, TabuLa, and other small language models, and highlighting their training strategies and conditioning mechanisms. The performance of each model across multiple datasets is evaluated based on 3 main criteria: downstream machine learning utility, data representativeness, and privacy preservation. Furthermore, we examine the potential of smaller language models to achieve competitive performance with significantly reduced computational cost. Experimental results show necessary trade-offs among model size, data quality, and efficiency, underscoring that larger models are not necessarily optimal.

Navigating dynamic environments is a fundamental challenge in distributed robotic systems, particularly when faults occur within the system itself, resulting in a changing connectivity graph. Classical graph search algorithms such as A* provide optimal paths as long as the graph is static. However, faults are a part of real life applications and cannot be ignored; classical approaches scale poorly in such scenarios because updating the graph topology requires extensive inter-robot communication, recombination of local maps, and replanning. This paper proposes a reinforcement learning (RL)-based approach that enables agents to learn navigation policies without requiring global knowledge of the graph. Each agent observes only its immediate neighborhood, making locally reasonable decisions about navigating toward a target location that collectively achieve near-optimal global performance. Through training on randomly chosen faults, our model learns robust traversal behaviors that adapt online to topology changes, reducing communication overhead compared to a more basic A*-based approach in a faulty environment.

This paper studies coordinated trajectory planning and tracking control for multiple unmanned surface vessels (USVs) under strict privacy requirements. To avoid the privacy risks associated with direct position sharing in conventional cooperative methods, the proposed approach adopts an estimated fleet centroid as the only shared variable, preventing individual trajectory disclosure while enabling coordination. Based on this interaction mechanism, a formation-oriented trajectory is generated for the fleet. The collective dynamics are modeled using Port-Hamiltonian systems, and a passivity-based tracking controller is designed for each USV to accurately follow the planned trajectories. The stability of the closed-loop system is rigorously proven, and experiments on a real USV platform confirm effective formation tracking and privacy preservation. The proposed result extends and validates through experimental results the approach in [26] that was limited to idealized pointmass models and lacked a feedback control.