Advanced intelligent systems (Weinheim an der Bergstrasse, Germany)最新文献

A key goal in evacuation management is to quickly and safely remove panicking crowds from buildings, festivals, or airplanes while preventing crush fatalities. Recently, there has been much progress in realistically modeling crowds in complex environments, based on social force models, cellular automata, and machine learning. However, current models assume specific social interactions and do not allow to systematically explore how to optimize crowd cooperation and evacuation. In contrast, the present work focuses on the question, how an ideal crowd of superintelligent agents, comprising humans, robots, or smart active particles, would cooperate to optimize evacuation. A method is developed that uses multiagent reinforcement learning combined with self-play to learn optimal crowd behavior from scratch. Crucially, the agents in this approach are pressure-aware and autonomously learn collision and crushing avoidance. After training, they adopt interpretable evacuation strategies like queuing and zipper merging and outperform traditional evacuation models in terms of fatality avoidance and evacuation rate. Our method can be used to enhance guidelines for mass evacuation, potentially saving lives.

Infertility has emerged as a significant health issue impacting individuals’ lives. In prior investigations, image classification has been applied to identify morphologic abnormalities associated with infertility issues. However, the limited data availability has impeded high performance. In the field of image augmentation techniques, particularly concerning generative adversarial networks (GANs), an alternative approach can encounter a significant issue known as mode collapse. This phenomenon arises when the generator consistently produces a restricted set of identical or highly similar images, which may negatively affect the overall performance and accuracy of the model. Consequently, the aim of this study is to mitigate mode collapse by employing loss-based ensemble GAN framework, formulated based on the integration of two distinct GAN models. In addition, a comprehensive analysis is carried out using an expanded approach involving three GAN models in conjunction with a spatial augmentation technique. The Shifted Window Transformer model achieves 95.37% accuracy on the HuSHeM dataset, outperforming other classification models. This finding shows enhanced accuracy relative to earlier studies using the identical dataset.

Since ancient times, oracles (e.g., Delphi) has the ability to provide useful visions of where the society is headed, based on key event correlations and educated guesses. Currently, foundation models are able to distill and analyze enormous text-based data that can be used to understand where societal components are headed in the future. This work investigates the use of three large language models (LLM) and their ability to aid the research of nuclear materials. Using a large dataset of Journal of Nuclear Materials papers spanning from 2001 to 2021, models are evaluated and compared with perplexity, similarity of output, and knowledge graph metrics such as shortest path length. Models are compared to the highest performer, OpenAI's GPT-3.5. LLM-generated knowledge graphs with more than 2 × 10⁵ nodes and 3.3 × 10⁵ links are analyzed per publication year, and temporal tracking leads to the identification of criteria for publication innovation, controversy, influence, and future research trends.

This article presents a computational method for generating virtual 3D morphologies of functional materials using low-parametric stochastic geometry models, that is, digital twins, calibrated with 2D microscopy images. These digital twins allow systematic parameter variations to simulate various morphologies, which can be deployed for virtual materials testing by means of spatially resolved numerical simulations of macroscopic properties. Generative adversarial networks (GANs) have gained popularity for calibrating models to generate realistic 3D morphologies. However, GANs often comprise numerous uninterpretable parameters, making systematic variation of morphologies for virtual materials testing challenging. In contrast, low-parametric stochastic geometry models (e.g., based on Gaussian random fields) enable targeted variation but may struggle to mimic complex morphologies. Combining GANs with advanced stochastic geometry models (e.g., excursion sets of more general random fields) addresses these limitations, allowing model calibration solely from 2D image data. This approach is demonstrated by generating digital twins for the morphology of microstructures in all-solid-state battery (ASSB) cathodes. Since the digital twins are parametric, they support systematic exploration of structural scenarios and their macroscopic properties. The proposed method facilitates simulation studies for optimizing 3D morphologies, benefiting not only ASSB cathodes but also other materials with similar structures.

Multisensory integration (MSI) is a core brain function underlying perception, learning, and behavior. Understanding the computational mechanisms of MSI is key to advancing AI and brain-inspired systems. While earlier models relied on probabilistic frameworks, recurrent neural networks (RNNs) offer advantages in capturing temporal dynamics and neural computations. This review presents a critical examination of computational models of MSI, focusing on the evolution from probabilistic integration to modern RNN-based methods. Biological evidence for temporal coordination in multisensory areas is analyzed and explored how different RNN architectures (e.g., vanilla, long short-term memory, and gated recurrent unit) simulate these dynamics. Comparative analyses show RNNs’ superiority in robustness and learning efficiency, with up to 46.9% improvement in classification tasks involving sensory fusion. We introduce a taxonomy of MSI tasks and a novel evaluation framework for model benchmarking. Real-world case studies—from speech recognition to prosthetic control—highlight practical applications. Challenges in interpretability, data efficiency, and generalization are also discussed. The review provides actionable insights for future research in both computational neuroscience and artificial intelligence. By bridging neurobiological principles and machine learning, RNN-based models pave the way for intelligent systems capable of flexible, context-aware multisensory processing.

The treatment plan for brain tumors varies depending on the type and stage of the tumor. Early diagnosis plays a vital role in determining appropriate treatment. In addition to clinical routines, artificial intelligence-based systems that produce automated, quantitative, and objective results can assist clinicians and scientists in making early diagnoses. For this motivation, this study proposes a deep learning-based system that classifies brain tumors obtained by magnetic resonance imaging. In the proposed approach, several wavelet transform approaches are applied to the raw dataset images. Thus, in addition to automated feature extraction in deep learning, it aimed to detect more detailed features. Therefore, four types of datasets have been obtained. Then, using the transfer learning approach, some popular convolutional neural network and vision transformer models are trained separately with the four-type datasets, and the test results are compared. The networks that produced the highest results are used to make the final decision with the ensemble technique. In the first analysis, the best performance was obtained using original data with an 83.50% accuracy value, and the second highest performance is obtained 81.72% accuracy value using the Daubhecies wavelet before deep learning. The third and fourth high performances are 81.47% and 81.22% accuracy, respectively, using original data. In the ensemble analysis, the highest result is achieved at 85.03% accuracy value using the bagging-ensemble approach of the networks, namely MobileNet-v3, vision transformer, ResNeXt, and DenseNet-201. This study demonstrates that using a hybrid wavelet transform and deep learning approach improves classification performance. This may inspire the use of the same method to solve different classification problems.

This study introduces a minimally invasive robotic probe inspired by plant root growth, designed for subsoil exploration and future ecosystem monitoring and intervention. The bio-inspired probe advances in soil by mimicking plant root apical growth, creating and consolidating a borehole through the injection of a bio-based, biodegradable binder at its tip. This innovative process confines penetration resistance to the tip while generating a hollow tubular structure by harnessing in situ local soil. The probe's penetration is facilitated by a linear actuator, which can be retracted upon reaching a desired depth, thereby minimizing the environmental dispersion of mechatronic components. This approach not only enhances the efficiency of subsoil exploration (whether on-Earth or in outer space) by reducing penetration force requirements and reliance on exogenous material but also ensures environmental sustainability by employing biodegradable materials and lowering mechanical footprints. The robotic probe's design and functionality highlight the potential of bio-inspired technologies to address complex environmental challenges, paving the way for future innovations in ecological research and conservation efforts. This study underscores the importance of integrating biological principles into engineering solutions to develop tools that are both effective and environmentally responsible.

This article presents a novel, versatile wearable force myography (FMG) system based on optical fiber technology, designed for high sensitivity and mechanical robustness. Unlike conventional FMG systems, which are susceptible to environmental interference, the proposed system utilizes light loss through controlled fiber–polymer contact to achieve stable and noise-free signal transmission. Its compact and flexible form factor allows seamless integration into wearable devices, facilitating muscle-activity monitoring under diverse real-world conditions, including biologically challenging scenarios such as sweating. Experimental evaluations highlight the system's ability to detect even micronewton-scale forces and accurately recognize multiple gestures. Furthermore, the system can estimate joint angles, including those of individual fingers, which underscores its potential for precise motion capturing and continuous tracking. Overall, the proposed FMG system represents a promising solution for a wide range of practical human–robot interaction applications.

Kidney stones can cause severe pain and complications like chronic kidney disease. Although retrograde intrarenal surgery with laser lithotripsy is effective, current retrieval methods are inefficient, typically capturing only 1–3 fragments per ureteroscope pass and requiring many passes for full clearance. A novel spinner device that enables ultra-efficient fragment removal through spinning-induced localized suction is introduced. It generates spiral and circulating flows to capture fragments from over 20 mm away, eliminating the need to chase them. Optimized via computational fluid dynamics and validated in vitro and ex vivo, the spinner retrieves ≈60 small (0.5–2 mm) or ≈15 larger (2–3 mm) fragments per pass. It demonstrates nearly 100% capture of 60 fragments in bench tests and removes 45 fragments in 4 s in a porcine kidney model. This technology markedly improves procedural efficiency by reducing operative time, increasing stone-free rates, and minimizing the number of ureteroscope passes.

Continuum manipulators can operate in complex environments where traditional rigid manipulators fail. However, the modeling of inverse kinematics remains challenging because of its inherent nonlinearities and various external conditions. This work proposes an online learning control framework with a data cache pool utilizing a constant-curvature model inspired neural network (CCMINN) model to obtain the inverse kinematics model of tendon-actuated continuum manipulators. The CCMINN model is a kind of geometric relationships-inspired neural network, which is inspired by the geometric relationships within the constant-curvature model. This model improves the ability of traditional fully connected neural network models on high convergence speed and precision through its constant-curvature inspiration layers. These layers embed geometry insights into the neural network structure rather than loss functions like physics-informed neural networks. The online learning framework enables CCMINN to maintain high control accuracy in a variety of external load scenarios. Experiments show average tracking errors of 1.4 mm, 1.38 mm, and 1.48 mm (0.7%, 0.64%, and 0.74% of the continuum manipulator length) in the free space, under constant and variable loading conditions, respectively. The results show that combining the fast-converging CCMINN with an online learning control framework enables high-precision and robust positioning control of continuum manipulators under various external payloads.