Martin Ziegler

Modulated magnetoelectric (ME) sensors enable the detection of low-frequency, small-amplitude magnetic fields. Their performance is governed by a high-dimensional parameter space that combines the nonlinear coupling of magnetic, electric, and mechanical properties with the system electronics and multiple operating conditions. This complexity makes many tasks challenging, such as application-specific sensor design, systematic performance analysis, and identifying ideal operating conditions. Here, we take the first step toward navigating this complexity by introducing AI-driven surrogate models that systematize operating point selection and characterization workflows. Three artificial neural networks are trained on measured data of a modulated ME sensor based on the ΔE effect. The networks predict sensor admittance, noise spectral density (NSD), and signal output based on fundamental operating parameters, such as the excitation amplitude, excitation frequency, and the magnetic bias flux density. The models achieve accurate performance with sub-percent errors for sensor admittance, percent-level error for the NSD, and a 90th percentile error with 5% for the signal model. Based on the trained surrogates, the limit of detection can be reproduced reliably over most of the operating regime. A data-efficiency analysis further examines how model performance depends on the amount of training data. Finally, the surrogate models are used to solve three different operating point optimization problems of increasing complexity. Using a ΔE-effect sensor as a representative example, this work demonstrates that data-driven surrogate modeling can support characterization and optimization of modulated ME sensors. The approach may be extended to other ME sensor concepts and ultimately to device and array optimization.

Resistive switching devices based on complex oxide heterostructures exhibit rich, history‐dependent electrical responses arising from polarity‐selective transport mechanisms, defect dynamics, and interfacial phenomena. Accurately capturing these nonlinear behaviors remains a major challenge for predictive device modeling and for the reliable design of large‐scale memristive systems. Here, we introduce a dual‐branch, flux‐controlled extended memristor model that provides a physically interpretable asymmetric conduction in multilayer Au/HfO _x /Al ₂ O ₃ /TiO ₂ /TiN devices. The model is calibrated and validated against experimental current–voltage characteristics measured on laboratory‐fabricated devices. The approach decomposes the device memductance into two polarity‐activated branches that reproduce the distinct forward and reverse transport regimes, while embedding flux as the internal state variable governing long‐term evolution and hysteresis shaping. A hybrid machine‐learning calibration pipeline—combining Latin Hypercube Sampling, Bayesian Optimization, and gradient‐based refinement—enables robust parameter identification directly from current–voltage experimental measurements, ensuring quantitative agreement across a broad range of time‐ and frequency‐dependent excitations. The resulting model captures key physical signatures such as lobe asymmetry, history dependence, and flux accumulation, providing a unified framework suitable for device characterization, circuit‐level simulation, and neuromorphic hardware design. Owing to its modular structure and physical consistency, the methodology can be readily extended to other classes of resistive switching materials and memory devices.

Reservoir computing (RC) is a powerful framework for predicting the temporal evolution of variables of nonlinear dynamical systems, yet the role of reservoir topology—particularly symmetry in connectivity and weights—remains not adequately understood. This work investigates how the structure of the reservoir network influences the performance of RC on time series prediction tasks derived from four dynamical systems of increasing complexity: the Mackey–Glass system with delayed-feedback, two nonlinear models of two-dimensional thermal convection flows, and a three-dimensional shear flow model exhibiting transition to turbulence. Using five reservoir topologies in which connectivity patterns and edge weights are controlled independently, we evaluate both direct- and cross-prediction tasks. The results show that symmetric reservoir networks improve prediction accuracy for the convection-based systems when the input dimension is smaller than the number of degrees of freedom. In contrast, the shear-flow model displays almost no sensitivity to topological symmetry due to its strongly chaotic high-dimensional dynamics. These findings reveal how structural properties of reservoir networks affect their ability to learn complex dynamics and provide guidance for designing more effective RC architectures.

The formation of extended sulfur vacancies in MoS ₂ monolayers is closely associated with catalytic activity and may also be the basis for its memristive behavior. Nanosecond‐scale molecular dynamics simulations using machine learning interatomic potentials (MLIPs) reveal key mechanisms of cooperative vacancy transport, including incorporation of vacancies into clusters of arbitrary size. The simulations provide a coherent atomistic explanation for irradiation‐induced vacancy patterns observed experimentally, especially the formation of line defects spanning tens of nanometers. Results and performance are compared of two MLIP frameworks: (i) on‐the‐fly learning with Gaussian approximation potential, and (ii) fine‐tuning of an equivariant foundation model.

Next-generation artificial intelligence (AI) hardware based on memristive devices offers a promising approach to reducing the increasingly large energy consumption of AI applications. However, programming memristive AI hardware to achieve a desired synaptic weight configuration remains challenging because it requires accurate and energy-efficient algorithms for selecting the optimal weight-update pulses. Here, we present a computationally efficient AI model for predicting the weight update of memristive devices and guiding device programming. The synaptic weight-update behavior of bilayer HfO2/TiO2 memristive devices is characterized over a range of pulse parameters to provide experimental data for the AI model. Three different artificial neural network (ANN) configurations are trained and evaluated regarding the amount of training data required for accurate predictions and the computational costs. Finally, we apply the model to an antipulse weight-update process to demonstrate its performance. The results show that accurate and computationally inexpensive predictions are possible with comparatively few datasets and small ANNs. The normalized weight-update processes are predicted with accuracies comparable with larger model architectures but require only 896 floating point operations and 8.33 nJ per inference. This makes the model a promising candidate for integration into AI-driven device controllers as a precise and energy-efficient solution for memristive device programming.

We show that connectivity within the high-dimensional recurrent layer of a reservoir network is crucial for its performance. To this end, we systematically investigate the impact of network connectivity on its performance, i.e., we examine the symmetry and structure of the reservoir in relation to its computational power. Reservoirs with random and asymmetric connections are found to perform better for an exemplary Mackey-Glass time series than all structured reservoirs, including biologically inspired connectivities, such as small-world topologies. This result is quantified by the information processing capacity of the different network topologies which becomes highest for asymmetric and randomly connected networks. Published by the American Physical Society 2025

Mankind is currently living in the era of mobile communication. Mobile communication encompasses almost all areas of our daily life and is heavily used in most sectors of economy, including agriculture, healthcare, education, and so on. With mobile devices such as smart phones, people can connect to any other people somewhere on the earth, can access huge databases via the internet, stream videos and movies, to name just a few of the many mobile services that are routinely used in the early 2020s. Artificial intelligence, virtual reality, and other emerging new applications will further boost the importance of mobile communication. It is a general trend that with progressing evolution of mobile communication, the amount of data to be transmitted wirelessly increases rapidly and continuously. To enable this, a suitable well-performing hardware infrastructure is needed. The hardware used so far in the successively introduced generations of systems for mobile communication is essentially based on fast transistors. We designate these transistors as gigahertz (GHz)–terahertz (THz) transistors, as the high data rates to be transferred require an electronic hardware operating properly at frequencies in the GHz–THz range. The present paper provides a comprehensive in-depth discussion of these transistors. After a short survey of the evolution of mobile communication systems, the different categories and classes of GHz–THz transistors are introduced, relevant transistor performance measures called figures of merit are defined, issues of transistor design are dealt with, and general design rules for GHz–THz transistors are established. This is followed by a tour through the evolution of these transistors, starting in the 1950s, when the first transistors with GHz capabilities were demonstrated, up to the latest developments since the turn of the millennium and to the state-of-the-art in mid-2024. The discussion includes evolutionary advances as well as true breakthroughs, both contributing to the significant performance improvements achieved over the years. Furthermore, some research activities on GHz–THz transistors, which, unfortunately, led to a dead end as well as ongoing efforts with still unclear outcome are presented and the approaching limits of transistor performance are critically examined. Finally, we take a short view on alternative hardware options for mobile communication systems beyond the limits of transistors.

Hopfield attractor networks are robust distributed models of human memory, but they lack a general mechanism for effecting state-dependent attractor transitions in response to input. We propose construction rules such that an attractor network may implement an arbitrary finite state machine (FSM), where states and stimuli are represented by high-dimensional random vectors and all state transitions are enacted by the attractor network’s dynamics. Numerical simulations show the capacity of the model, in terms of the maximum size of implementable FSM, to be linear in the size of the attractor network for dense bipolar state vectors and approximately quadratic for sparse binary state vectors. We show that the model is robust to imprecise and noisy weights, and so a prime candidate for implementation with high-density but unreliable devices. By endowing attractor networks with the ability to emulate arbitrary FSMs, we propose a plausible path by which FSMs could exist as a distributed computational primitive in biological neural networks.

Hopfield attractor networks are robust distributed models of human memory, but lack a general mechanism for effecting state-dependent attractor transitions in response to input. We propose construction rules such that an attractor network may implement an arbitrary finite state machine (FSM), where states and stimuli are represented by high-dimensional random vectors, and all state transitions are enacted by the attractor network's dynamics. Numerical simulations show the capacity of the model, in terms of the maximum size of implementable FSM, to be linear in the size of the attractor network for dense bipolar state vectors, and approximately quadratic for sparse binary state vectors. We show that the model is robust to imprecise and noisy weights, and so a prime candidate for implementation with high-density but unreliable devices. By endowing attractor networks with the ability to emulate arbitrary FSMs, we propose a plausible path by which FSMs could exist as a distributed computational primitive in biological neural networks.