Publications

Bio-nanocomposites, produced from renewable resources, stand out as one of the promising next-generation material classes for both biomedical and environmental applications due to their biocompatible and biodegradable structures. However, the vast and complex design space of these materials, which depends on numerous variables including the type of biopolymer matrix and the type and concentration of the nanostructured filler, makes their optimization quite challenging with traditional “trial-and-error” based research methods. This study provides a general overview of an artificial intelligence (AI)-based methodology that holds the potential to overcome this complexity. In this context, the limitations of the traditional “forward screening” approach are discussed, and the “inverse design” paradigm, based on the idea of “designing a material with desired properties,” is proposed as a solution. The study outlines the potential roles of various AI tools in this field, from supervised learning models to deep learning approaches and optimisation algorithms. This theoretical framework is substantiated by practical case studies that range from drug delivery to sustainable packaging. The case studies demonstrate how AI not only enables the prediction of properties but also facilitates strategically developing materials for a particular function, thereby accelerating the discovery process. Finally, current challenges, such as data scarcity and the interpretability of models, are addressed, and future visions to overcome these obstacles, including physics-informed AI, autonomous discovery platforms, and human-AI partnerships, are discussed. It is concluded that this synergy has the capacity to stimulate a new era of innovation in sustainable materials science.

Fuel desulfurization is crucial due to the harmful effects of acid rain, air pollution, and health issues associated with sulfur compounds from fossil fuel combustion. Stricter international regulations require reducing sulfur content in fuels to ultra-low levels (<15 ppm), which current hydrodesulfurization (HDS) technology struggles to achieve due to high energy costs and inefficiencies with refractory aromatic sulfur compounds, such as dibenzothiophene (DBT). This gap in technology is driving the development of new solutions. This study explores the integration of materials science, chemical engineering, and artificial intelligence (AI) to meet ultra-low sulfur targets. It highlights alternative strategies, such as adsorptive (ADS) and oxidative (ODS) desulfurization, using advanced nanocomposite materials, including Metal-Organic Frameworks (MOFs), graphene hybrids, and zeolites. These innovative materials focus on “precision targeting” and “molecular recognition,” offering greater efficiency under milder conditions compared to traditional HDS methods AI and computational science are crucial for identifying the most suitable materials among a vast array of candidates. Combining Density Functional Theory (DFT) with Machine Learning (ML) accelerates material discovery by transitioning from a trial-and-error approach to an “inverse design” method. Successful AI tools, including Artificial Neural Networks (ANNs) and Genetic Algorithms (GAs), have been proven effective in optimising desulfurization processes through real-world examples. In summary, achieving cleaner fuels requires innovative materials guided by AI. The study also addresses the challenges of scaling this interdisciplinary approach and envisions the potential for transforming desulfurization from a cost centre into a profit centre through precise molecular separation.

With the proliferation of devices on the Internet of Things (IoT), ensuring their security has become paramount. Device identification (DI), which distinguishes IoT devices based on their traffic patterns, plays a crucial role in both differentiating devices and identifying vulnerable ones, closing a serious security gap. However, existing approaches to DI that build machine learning models often overlook the challenge of model generalizability across diverse network environments. In this study, we propose a novel framework to address this limitation and to evaluate the generalizability of DI models across data sets collected within different network environments. Our approach involves a two-step process: first, we develop a feature and model selection method that is more robust to generalization issues by using a genetic algorithm with external feedback and datasets from distinct environments to refine the selections. Second, the resulting DI models are then tested on further independent datasets to robustly assess their generalizability. We demonstrate the effectiveness of our method by empirically comparing it to alternatives, highlighting how fundamental limitations of commonly employed techniques such as sliding window and flow statistics limit their generalizability. Moreover, we show that statistical methods, widely used in the literature, are unreliable for device identification due to their dependence on network-specific characteristics rather than device-intrinsic properties, challenging the validity of a significant portion of existing research. Our findings advance research in IoT security and device identification, offering insight into improving model effectiveness and mitigating risks in IoT networks.

Introduction Semantic verbal fluency (SVF) is a widely used measure of frontal executive function and access to semantic memory. SVF scoring metrics include the number of unique words generated, perseverations, intrusions, semantic cluster size and switching between clusters, and scores vary depending on the language the test is administered in. In this paper, we review the existing normative data for Turkish, the main metrics used for scoring SVF data in Turkish, and the most frequently used categories. Method We conducted a systematic review of peer-reviewed papers using Medline, EMBASE, PsycInfo, Web of Science, and two Turkish databases, TR-Dizin and Yok-Tez. Included papers contained data on the SVF performance of healthy adult native speakers of Turkish, and reported the categories used. Versions of the SVF that required participants to alternate categories were excluded. We extracted and tabulated demographics, descriptions of groups, metrics used, categories used, and sources of normative data. Studies were assessed for level of detail in reporting findings. Results 1400 studies were retrieved. After deduplication, abstract, full text screening, and merging of theses with their published versions, 121 studies were included. 114 studies used the semantic category “animal”, followed by first names (N = 14, 12%). All studies reported word count. More complex measures were rare (perseverations: N = 12, 10%, clustering and switching: N = 5, 4%). Four of seven normative studies reported only word count, two also measured perseverations, and one reported category violations and perseverations. Two normative studies were published in English. Conclusions There is a lack of normative Turkish SVF data with more complex metrics, such as clustering and switching, and a lack of normative data published in English. Given the size of the Turkish diaspora, normative SVF data should include monolingual and bilingual speakers. Limitations include a restriction to key English and Turkish databases.

Smart speakers in multi-user spaces, such as Amazon Echos, introduce risks to both owners and anyone sharing the space. They store voice recordings of user requests, and anyone in range can potentially interact with the device. As smart speakers are usually bound to a single account, despite being shareable by design, it introduces potential tensions between users. We systematically searched the literature for findings on concerns and scenarios in which problems may arise and synthesised the resulting 20 papers in a narrative review. Owners were concerned about other users’, potentially malicious, interactions, device faults, and third party sharing. In contrast, bystanders worried about "being listened" to and a lack of awareness and protections. Our findings show a clear gap in literature on the privacy concerns of regular and incidental secondary users of smart speakers.

In recent years WiFi became the primary source of information to locate a person or device indoor. Collecting RSSI values as reference measurements with known positions, known as WiFi fingerprinting, is commonly used in various positioning methods and algorithms that appear in literature. However, measuring the spatial distance between given set of WiFi fingerprints is heavily affected by the selection of the signal distance function used to model signal space as geospatial distance. In this study, the authors proposed utilization of machine learning to improve the estimation of geospatial distance between fingerprints. This research examined data collected from 13 different open datasets to provide a broad representation aiming for general model that can be used in any indoor environment. The proposed novel approach extracted data features by examining a set of commonly used signal distance metrics via feature selection process that includes feature analysis and genetic algorithm. To demonstrate that the output of this research is venue independent, all models were tested on datasets previously excluded during the training and validation phase. Finally, various machine learning algorithms were compared using wide variety of evaluation metrics including ability to scale out the test bed to real world unsolicited datasets.