Abstract

Abstract
Driving simulators are essential tools for training, education, research, and scientific experimentation. However, the diversity and quality of virtual environments in simulations is limited by the specialized human resources availability for authoring the content, leading to repetitive scenarios and low complexity of real-world scenes. This work introduces a pipeline that can process text-based narratives outlining driving experiments to procedurally generate dynamic traffic simulation scenarios. The solution uses Retrieval-Augmented Generation alongside local open-source Large Language Models to analyse unstructured textual information and produce a knowledge graph that encapsulates the world scene described in the experiment. Additionally, a context-based formal grammar is generated through inverse procedural modelling, reflecting the game mechanics related to the interactions among the world entities in the virtual environment supported by CARLA driving simulator. The proposed pipeline aims to simplify the generation of virtual environments for traffic simulation based on descriptions from scientific experiment, even for users without expertise in computer graphics.

Abstract
Data-based approach for diagnosis of thyroid disorders is still at its early stage. Most of the research outcomes deal with binary classification of the disorders, i.e. presence or not of some pathology (cancer, hyperthyroidism, hypothyroidism, etc.). In this paper we explore deep learning (DL) models to improve the multi-class diagnosis of thyroid disorders, namely hypothyroid, hyperthyroid and no pathology thyroid. The proposed DL models, including DNN, CNN, LSTM, and a hybrid CNN-LSTM architecture, are inspired by state-of-the-art work and demonstrate superior performance, largely due to careful feature selection and the application of SMOTE for class balancing prior to model training. Our experiments show that the CNN-LSTM model achieved the highest overall accuracy of 99%, with precision, recall, and F1-scores all exceeding 92% across the three classes. The use of SMOTE for class balancing improved most of the model’s performance. These results indicate that the proposed DL models not only effectively distinguish between different thyroid conditions but also hold promise for practical implementation in clinical settings, potentially supporting healthcare professionals in more accurate and efficient diagnosis. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.

Abstract
The Vision Transformer (ViT) architecture has emerged as a potential game-changer in computer vision, offering scalability and global attention that have generated considerable interest in recent years. Its adaptability has fueled enthusiasm for its application. This work investigates the boundaries of the architecture, focusing on developing new techniques targeting explicitly complex tasks, such as medical imaging datasets, which often exhibit high variability, class imbalance, and limited sample sizes. We propose a set of mixed regularisation and augmentation techniques to enhance the performance of models. These include a novel loss function and a smoothly differentiable activation function, leading to more stable training and model performance. The results show that incorporating these techniques improves model performance and training convergence.

Abstract
Stress in the workplace is a major problem that affects people of all ages, backgrounds, and occupations. It can contribute to various health problems, from anxiety to insomnia, among others. Workplace stress significantly impacts employee well-being and productivity. Current stress-management approaches, while valuable, primarily address stress after it has occurred. This highlights the critical need for proactive systems capable of anticipating individual stress and preventing negative health consequences. This research presents the design and initial implementation of a novel microservice-based recommendation system for proactively mitigating workplace stress among computer users. The system leverages predicted stress levels to deliver timely, personalized, and easily implemented interventions. This study focuses on evaluating the system's architecture, core functionalities, and initial performance using a content-based filtering approach. A pilot study demonstrated the system's feasibility, highlighting areas for future development.

Abstract
Distributed acoustic sensing (DAS) is an instrumental approach that allows fiber optic cables to be turned into dense arrays of acoustic sensors. This technology, based on an optical time-domain reflectometer (OTDR) technique, is attractive in marine environments where instrumentation is difficult to implement. Its main applications lie in seismology, oceanography, and bioacoustics.Current technology limits the range of DAS to ca. 150 km making it very useful in the Azores, where seismic stations only exist on the Islands with a strong E-W alignment, as shown by Matias et al. (2021). The Azores have been suffering an increase in extreme wave conditions that affect navigation and coastal infrastructures. DAS can provide proxies for significant wave height, period, and surface currents on the shallow sections of the cable, complementing existing monitoring networks.The Azores region is part of the migration routes for fin and blue whales, which are known to produce acoustic signals during certain parts of the year. These vocalizations provide crucial data for Passive Acoustic Monitoring that can be used to support the establishment and update of mitigation measures addressing their preservation. DAS has already demonstrated its usefulness in detecting and tracking baleen whales using acoustic records.One issue that needs to be addressed in using DAS data is calibration. It is well demonstrated that strain or strain rate as measured by DAS can be converted to ground motion along the direction of the submarine cable section, if the apparent phase velocity is known. Similarity between DAS converted signals and co-located seismograms is well demonstrated but the absolute value is likely to vary with the cable coupling to the seafloor.This work reports on the recent operation of a DAS interrogator deployed at the Faial landing site to monitor the first 50 km of the telecommunication cable between Faial and Flores islands operated by Fibroglobal. The instrument used, HDAS developed by the IO-CSIC, recorded at 50 Hz for 11 months starting on the 15th of December 2023 with 10 m gauge length. For calibration purposes two 4C OBS were deployed close to the cable at ~10 km and ~30 km distance from the landing point. The OBSs were deployed in July 2024 and recovered in November 2024, providing 5 months of simultaneous recordings with the DAS.As expected, both earthquakes and whale vocalizations were identified on the DAS and OBS. We show that DAS can contribute to an improved localization of local offshore earthquake parameters due to its high density of sensors, particularly for the events occurring NW of Faial Island, with locations North of the cable. Clear landward and seaward oceanic waves are identified on the cable's shallow section. In all the applications the main question to address is the variable coupling of the cable to the seafloor in the Azores plateau of volcanic origin.This work is supported by the Portuguese Fundação para a Ciência e Tecnologia, FCT, I.P./MCTES through national funds (PIDDAC): UID/50019/2025 and LA/P/0068/2020 (https://doi.org/10.54499/LA/P/0068/2020), by the MODAS project 2022.02359.PTDC, and by EC project SUBMERSE project HORIZON-INFRA-2022-TECH-01-101095055.