Application Of Neural Networks Research Articles

Reconfigurable optoelectronic neuromorphic properties of MoS2-based memristors decorated with Pt nanoparticles for low power spiking neural network applications

Abstract The intriguing properties of two-dimensional (2D) materials render them attractive for energy efficient neuromorphic computations because their atomic scale thickness can alleviate the power requirements of the device. In parallel, their layered structure can be leveraged to optically program the device and further reduce power consumption. Along these lines, in this work, a forming free memory device consisting of a MoS2 monolayer with dimensions of ~100 μm decorated with small (~ 3 nm diameter) Pt nanoparticles (NPs) was fabricated. The impact of the Pt NPs’ surface density on the optoelectronic neuromorphic properties under ultraviolet irradiation (λ = 390 nm) was systematically investigated. More specifically, the reference samples without Pt NPs exhibited only synaptic behavior, while the NPs-based one (surface density of ~ 2 x 1012 NPs/cm2) presented a neuron-like response. An elevated surface density (~ 5 x 1012 NPs/cm2) just reduced the frequency of the generated spikes. Various synaptic plasticity and neuronal coding schemes were experimentally demonstrated. The underlying origins of this behavior were attributed to band bending in the Pt NPs-MoS2 interface, leading to a trapping effect of electrons on the metallic nanoparticles, evidenced by photoluminescence quenching, followed by a detrapping process, demonstrated by the reduced firing rate when using the Pt layer with the higher surface density. This versatility of the devices was leveraged to simulate the behavior of a fully optoelectronic spiking neural network. Considering the low energy consumption per spike (~400 pJ) during the experimental recorded neuromorphic properties, a dramatically reduced power consumption of ~320 μW was extracted during the pattern recognition of optical images. Our work provides valuable insights for emulating the artificial synaptic and neuronal behavior and paves the way for the development of next-generation and fully memristive artificial neural networks with a very small energy footprint.

On‐Chip Optoelectronic Synapse Based on Embedded Amorphous Carbon in Lithium Niobate‐on‐Insulator

AbstractLithium niobate on insulator (LNOI) wafers serve as a fundamental photonic platform for photonic integrated circuits. However, their application in photonic neural networks is limited due to a lack of optoelectronic properties, necessitating complex heterogeneous integration. This work proposes a novel integration scheme that embeds heterogeneous structures within LNOI using ion implantation to achieve broadband optoelectronic functionalities (photovoltaic and persistent photo‐conductivity) for mimcking synaptic functions for Pavlovian‐type associative learning. Implanted carbon ions form an AC layer embedded in a lithium niobate layer (AC – LNOI), characterized by conductivity and wide‐spectrum light absorption. Upon illumination, the AC generates photo carriers, endowing AC‐LNOI with the persistent photo‐conductivity, which is employed as an artificial synapse to mimic synaptic functionalities. This work demonstrates the embedded heterogeneous integration as an efficient avenue to reduce the complexity within artificial neural network learning systems based on photonic integrated circuits.

From Iterative Methods to Neural Networks: Complex-Valued Approaches in Medical Image Reconstruction

Complex-valued neural networks have emerged as an effective instrument in image reconstruction, exhibiting significant advancements compared to conventional techniques. This study introduces an innovative methodology to tackle the difficulties related to image reconstruction within medical microwave imaging. Initially, in the estimation phase, the proposed methodology integrates the Born iterative method with quadratic programming. Subsequently, in the refinement stage, the study explores the application of complex-valued neural networks to enhance the quality of reconstructions. The research emphasizes distinct complex-valued neural network architectures, namely, CV-UNET, CV-CNN, CV-MLP, and their corresponding performances. CV-UNET stands out as the best architecture, surpassing conventional methods and the other complex-valued neural networks variants. The complex-valued neural network improves the fidelity of reconstructions and simplifies the procedure by obviating the need for multiple training steps, a common prerequisite in real-valued neural networks.

Data Analytics by Python AI, to Understand Deviation of Minimum Temperature During Winter (Using Artificial Neural Network)

This paper is based on data analysis by Python programming language on Google collaborator platform, supported by the AI code, model used as neural network, to understand the deviation of minimum temperature for the period November to February in comparison with recorded normal. The source of data is online data collection platform ‘INDIA METEOROLOGICAL DEPARTMENT, PUNE’ and the online site ‘OGIMET’. The surface data for Alipore (42807) and the recorded normal temperature, as available there has been collected in csv file format, then uploading the csv file in Google collaborator platform, executed analysis by Python neural network model technique, to understand the predicted output, i.e. pattern of deviation of minimum temperature, based on analysis of big data, with minimum temperature data from 1969 to 2024. In this case the input column used is the column created with data of difference between minimum temperature and recorded normal for that month. The mod value of temperature difference for the four months, November to February filtered from this big data set has been used as filtered and scaled data, subjected to analyse to understand the future trend of this deviation. Model fit has been used with train-test split in 80%-20% ratio. The artificial neural network model actually resembles the structure of human brain, which can perform intelligent tasks similar to human brain by the process of learning by the machine. Like human brain a similar network is there with mapping between input and output, information propagated through different layers, using back propagation technique to learn the model to get best output with minimum loss. This technique resembles to the networking of human brain and neuron, where information is propagated through synaptic joints to react accordingly. In this paper prediction as well as analysis has been made with the application of artificial neural network with used model as LSTM supported by activation function and optimiser to learn the model for best fit output. The loss value of the model has also been verified to understand the success rate of the model

Application of Convolutional Neural Networks for Discriminating Anthropogenic and Natural Earthquakes in Madagascar

ABSTRACT This study uses a convolutional neural network (CNN) to classify waveforms into distinct categories of events, such as earthquakes and blasts (including quarry and mining explosions), using spectrograms, with a specific application in Madagascar. This approach consists of three key steps: (1) generating spectrograms from ground-motion recordings to represent time–frequency information; (2) training, validating, and testing the model using the generated spectrograms; and (3) making predictions. The performance of the CNN model is evaluated using a commonly used loss function and accuracy measures. This study utilizes data from seven short-period permanent stations located in the central part of Madagascar. Initially, we use part of the data for training, validation, and testing, focusing on 2492 known events. The results indicate that despite a limited training dataset and a few stations, the testing process achieves an average model accuracy of 0.98. The mean precision for each class is 0.91 for quarries and 0.99 for earthquakes, with mean recall values of 0.94 for quarries and 0.99 for earthquakes. The F1-scores are 0.92 for quarry blasts and 0.99 for earthquakes. The derived CNN model is subsequently adopted to predict the remaining uncategorized events, with predictions made on a station-by-station basis to ensure better results. The results demonstrate that the model successfully identifies mining sites, even those that were not included in the training dataset. Therefore, the model has effectively learned the distinguishing features of different event classes, even when presented with data from various parts of Madagascar. This work underscores the importance of machine learning techniques in seismology and highlights the potential of CNNs to improve seismic event discrimination. The performance of the present study is further assessed by comparing our results with existing CNN methods utilizing key metrics such as accuracy, precision, recall, and F1-score.

First demonstration of leaky-integrate and fire neuron based GAA nanosheet FET with ultra-low energy consumption down to 0.8fj/spike for spiking neural network applications

Abstract The fundamental component of an artificial spiking neural network (SNN) is an electronic device designed to emulate a biological neuron effectively. However, a key concern is the high energy consumption and large area associated with these artificial neurons, rendering them highly inefficient. This article presents a CMOS-compatible Gate-All-Around Nanosheet Field-Effect Transistor (GAA NSFET)-based Leaky Integrate-and-Fire (LIF) neuron with gate length of 50nm. This design achieves an exceptionally low energy consumption of 0.8 fJ, thereby establishing a new benchmark in the field. Utilizing well-calibrated 3D TCAD simulation and the SRH recombination model, along with the Unibo2 model to represent impact ionization phenomena, the proposed GAA NSFET LIF neuron effectively captures integration and recombination aspects within it. Moreover, it operates with a supply voltage of merely 1.8V, significantly lower than conventional bulk FinFET and PD-SOI MOSFET-based LIF neurons. Also, it demonstrates spiking frequency of ~100MHz, approximately 10^5 times higher than that of a biological neuron, controlled by the input voltage of the device, allowing the neuron adapt dynamically. Nanosheet FET technology offers superior electrostatic control, minimizing parasitic capacitances and facilitating rapid switching capabilities, thereby optimizing both power and performance simultaneously. Furthermore, the effective area of the NSFET amounts to 0.004 μm^2, making it an appealing choice for low-power, area-efficient, large-scale hardware implementations of SNN applications.

AI in Risk Assessment and Management in Finance

The main contribution of core financial discipline, for instance, risk assessment and management to financial stability is for example to determine decisions, credit evaluation and fraud detection in the investment. As application of advanced artificial intelligence (AI) for achieving higher accuracy and efficiency in risk prediction in the financial market demand is growing, it is needed to work on it. In this research, we used the application of deep neural networks as an ‘advanced AI’ such that it would help manage and assess financial risk. Another application of DNNs is its use in the prediction process in credit scoring, portfolio optimization, and anomaly detection due to their good performance in identifying fine patterns in gigantic financial datasets leading to improved accuracy in the prediction. This research can be performed using the widely used AI tool TensorFlow as it is scalable, equipped with all the deep learning skills, and already implemented. Integration of TensorFlow to the operation of large scale data and of proactive detection of risks by financial institutions using Deep Neural Networks (DNNs). As a result, a gain in terms of classifications regulation compliance and a reduced loss of money are obtained by leveraging the AI based intelligence to boost watchfulness on the financial edge. AI’s impact on financial risk management is transformational, can be based on data, and can constitute a data driven, resiliant basis for financial strategies.

Characterization of subsurface anomalies in non-stationary thermal wave imaging using convolutional neural networks

Abstract An important development in the field of thermographic study is the application of convolutional neural networks (CNNs) to characterize subsurface abnormalities using quadratic frequency-modulated thermal wave imaging. This technology greatly improves the accuracy of identifying subsurface flaws by utilizing deep learning techniques, which makes it very useful for a variety of industrial applications cantered on material integrity evaluation. In this study, we present a new post-processing modality based on regression that is intended to efficiently characterize subsurface anomalies. Through experimental experiments involving carbon fibre reinforced polymers (CFRP) and glass fibre reinforced polymers (GFRP), our methodology has been rigorously confirmed. The purpose of the studies was to demonstrate the ability of our method to visualize subsurface structures, with a focus on precisely measuring their sizes. We evaluate our method's efficacy by contrasting the outcomes with current methodologies, employing signal-to-noise ratio (SNR) as a crucial performance parameter. This assessment proves our method's superiority in offering more trustworthy and lucid insights about underlying structures. In the end, our research adds to the expanding body of knowledge about non-destructive testing and has encouraging ramifications for material science and engineering applications in the future.

Orthogonal polynomial-based neural network solution for differential equations with implicit boundary conditions

Purpose This paper aims to provide the approximate solution to variety of differential equations with implicit boundary conditions. Specifically, the focus is on solving higher-order differential equations and system of differential equations, fluid mechanics problem using an orthogonal polynomial-based neural network with extreme learning machine (ELM) algorithm. Design/methodology/approach The authors use neural network constructed with four different types of orthogonal polynomials: Legendre polynomial, Laguerre polynomial, Chebyshev polynomial and Hermite polynomial and train it using ELM algorithm. The neural network consists of single hidden layer replaced by functional expansion block which uses orthogonal polynomial to extract its features. The result of neural network provides approximate solutions to the problems. To show the capability and efficacy of the approach, obtained solutions are compared with the exact solutions and those derived from traditional numerical technique. Findings Numerical and comparative studies show that the neural network solutions are obtained with better accuracy. Moreover, the used approach is simple to implement and offer robust framework for solving differential equations with complex boundary conditions. Originality/value Implicit boundary value problems are successfully addressed, yielding closed form solutions that are highly valuable for real-world applications. Further, the proposed approach for solving implicit problems enhances the applicability of orthogonal polynomial-based neural network with ELM.

Exoplanet Detection Using CNN

ABSTRACT Detection of exoplanets, which are planets orbiting stars other than our Sun, is very important in understanding planetary systems and their habitability. This research examines the applicability of Convolutional Neural Networks (CNNs) in the detection of exoplanets using light curve data obtained from space telescopes like Kepler and TESS. The main aim is to automate and improve the accuracy in the identification of exoplanetary transits in noisy datasets. A CNN-based model was developed and trained on a labeled dataset comprised of light curves, which were labeled with binary data that indicated the presence or absence of an exoplanet. The architecture was optimized for feature extraction from time-series data, capturing subtle variations in brightness indicative of planetary transits. The proposed CNN achieved 98% accuracy, significantly outperforming traditional methods such as manual vetting and classical machine learning models. Results show that the model is indeed robust in distinguishing transit signals from stellar variability and instrumental noise. The analysis further indicated that the model generalizes well to unseen datasets, reducing false positives and negatives significantly. This work concludes that CNNs are a powerful, scalable approach to exoplanet detection: faster and more reliable. Future work would be the extension of the model to multiple planet systems and the consideration of other astrophysical parameters for greater precision. Keywords: Kepler, TESS , Planetary transits , Stellar variability , Habitability.

Integration of Grab Methods with Artificial Neural Networks for Enhanced Decision-Making Systems

This study explores the architecture, development, and practical applications of artificial neural networks (ANNs), with advances in Gray Relational Analysis (GRA) techniques. ANNs are computational models designed to model biological neural systems, consisting of layers of interconnected neurons—i.e., input, hidden, and output layers—connected by synaptic weights. Operating through a connectionist approach, these networks effectively mimic the four basic functions of biological neurons: receiving input, integrating information, processing data, and generating output. Traditional ANNs, although powerful, face limitations in embedded systems due to their reliance on high-precision digital information transfer and the resulting resource demands. This has prompted the development of more efficient alternatives, such as spiking neural networks (SNNs), which use event-driven spiking signals to reduce power consumption and memory usage. The field has advanced further by incorporating theoretical models for quantum neural computing and the use of genetic algorithms employing various crossovers and mutation techniques. The versatility of ANNs is evident in a variety of applications. In healthcare, they aid in pattern recognition associated with conditions such as breast cancer and diabetes. For water resource management, ANN models predict relationships between rainfall and water levels. In financial sectors, they analyze complex economic conditions and assess credit risk for small business loans. Industrial applications include modelling complex systems in manufacturing plants, although widespread adoption in this field is limited. Complementing neural network advances, GRA methods have emerged to address multi-criteria decision-making challenges. Notable advances include the GRAS techniques have been developed to correct matrices with negative values, while fuzzy GRA methods now include interval-valued triangular fuzzy numbers and probabilistically uncertain linguistic word sets. Recent advances include innovations such as score values and normalized Hamming distances within single-valued neutrosophic fuzzy stein summary, these computational approaches offer powerful ways to solve complex, multidimensional problems, with recent studies highlighting improved performance and promising prospects for future application.

Integration of Grab Methods with Artificial Neural Networks for Enhanced Decision-Making Systems

This study explores the architecture, development, and practical applications of artificial neural networks (ANNs), with advances in Gray Relational Analysis (GRA) techniques. ANNs are computational models designed to model biological neural systems, consisting of layers of interconnected neurons—i.e., input, hidden, and output layers—connected by synaptic weights. Operating through a connectionist approach, these networks effectively mimic the four basic functions of biological neurons: receiving input, integrating information, processing data, and generating output. Traditional ANNs, although powerful, face limitations in embedded systems due to their reliance on high-precision digital information transfer and the resulting resource demands. This has prompted the development of more efficient alternatives, such as spiking neural networks (SNNs), which use event-driven spiking signals to reduce power consumption and memory usage. The field has advanced further by incorporating theoretical models for quantum neural computing and the use of genetic algorithms employing various crossovers and mutation techniques. The versatility of ANNs is evident in a variety of applications. In healthcare, they aid in pattern recognition associated with conditions such as breast cancer and diabetes. For water resource management, ANN models predict relationships between rainfall and water levels. In financial sectors, they analyze complex economic conditions and assess credit risk for small business loans. Industrial applications include modelling complex systems in manufacturing plants, although widespread adoption in this field is limited. Complementing neural network advances, GRA methods have emerged to address multi-criteria decision-making challenges. Notable advances include the GRAS techniques have been developed to correct matrices with negative values, while fuzzy GRA methods now include interval-valued triangular fuzzy numbers and probabilistically uncertain linguistic word sets. Recent advances include innovations such as score values and normalized Hamming distances within single-valued neutrosophic fuzzy stein summary, these computational approaches offer powerful ways to solve complex, multidimensional problems, with recent studies highlighting improved performance and promising prospects for future application.

Мониторинг деформаций перегонных туннелей метрополитена с применением нейросетевого подхода

The article discusses application of artificial neural networks and machine learning algorithms for monitoring of deformations in subway tunnels located in complex mining and geological conditions. Particular attention is given to industrial and environmental safety, as well as modern methods for measuring crustal deformations using GPS/GLONASS technologies, geodetic, and mine surveying. The main stages of artificial neural networks operation are described, i.e. the training based on the tunnel parameters and conditions, testing, validation, and operation to predict the potential deformations. The key neural network architectures are considered such as the deep, convolutional, and recurrent networks along with their data processing capabilities. Examples are provided of artificial neural networks used for data interpolation, hazardous zone recognition, and tunnel ring monitoring. The importance of high-quality initial data, including geometric parameters, physical material properties, climatic conditions, and historical data, is emphasized. Implementation of artificial neural networks can help to promptly identify risks, predict the deformation dynamics, and classify the deformation types, enabling timely measures to prevent emergencies.

Разработка нейронных сетей для анализа вибрационных сигналов горного оборудования и предупреждения аварийных ситуаций

The article discusses the application of neural networks to analyze vibration signals of mining equipment in order to predict and prevent emergencies. A multilayer architecture of neural network trained has been developed using data from vibration sensors of real equipment. The proposed method allows to achieve high accuracy (95%) in classifying vibration patterns corresponding to different states of the equipment, i.e. from normal operation mode to pre-emergency states. Based on the analysis of spectral and temporal characteristics of vibrations, the method provides early (30-120 minutes) warning on development of potential faults. Experimental studies on real data confirm the ability of the neural network approach to reduce the number of emergency downtimes by 40% and the cost of emergency repairs by 25%. The proposed method opens prospects for creation of intelligent systems to ensure safety and efficiency of mining equipment based on predictive analytics.

MultiSEss: Automatic Sleep Staging Model Based on SE Attention Mechanism and State Space Model

Sleep occupies about one-third of human life and is crucial for health, but traditional sleep staging relies on experts manually performing polysomnography (PSG), a process that is time-consuming, labor-intensive, and susceptible to subjective differences between evaluators. With the development of deep learning technologies, particularly the application of convolutional neural networks and recurrent neural networks, significant progress has been made in automatic sleep staging. However, existing methods still face challenges in feature extraction and cross-modal data fusion. This paper introduces an innovative deep learning architecture, MultiSEss, aimed at solving key issues in automatic sleep stage classification. The MultiSEss architecture utilizes a multi-scale convolution module to capture signal features from different frequency bands and incorporates a Squeeze-and-Excitation attention mechanism to enhance the learning of channel feature weights. Furthermore, the architecture discards complex attention mechanisms or encoder–decoder structures in favor of a state–space sequence coupling module, which more accurately captures and integrates correlations between multi-modal data. Experiments show that MultiSEss achieved accuracy results of 83.84% and 82.30% in five-fold cross-subject testing on the Sleep-EDF-20 and Sleep-EDF-78 datasets. MultiSEss demonstrates its potential in improving sleep stage accuracy, which is significant for enhancing the diagnosis and treatment of sleep disorders.

A Systematic Review of the Application of Graph Neural Networks to Extract Candidate Genes and Biological Associations.

The development of high throughput technologies has resulted in the collection of large quantities of genomic and transcriptomic data. However, identifying disease-associated genes or networks from these data has remained an ongoing challenge. In recent years, graph neural networks (GNNs) have emerged as a promising analytical tool, but it is not well understood which characteristics of these models result in improved performance. We conducted a systematic search and review of publications that used GNNs to identify disease-associated biological interactions. Information was extracted about model characteristics and performance with the goal of examining the relationship between these factors and performance. Data leakage was found in 31% of these models. For node level tasks, univariate positive associations were identified between model accuracy and use of hyper parameter optimization, data leakage via hyperparameter optimization, test set size, and total dataset size. Among graph level tasks, an increase in AUC was identified in association with testing method and a decrease with optimization reporting. Data leakage may pose an issue for GNN-based approaches; the adoption of best practice guidelines and consistent reporting of model design would be beneficial for future studies.

Application of principal component analysis and Artificial neural networks for the prediction of QoS in FSO links over South Africa

Application of principal component analysis and Artificial neural networks for the prediction of QoS in FSO links over South Africa

RuptureNet2D, a Deep Neural Network Based Surrogate for Dynamic Earthquake Rupture Simulation in Two Dimensions

AbstractEarthquake dynamic rupture simulations are crucial for physics‐based seismic hazard assessment. However, due to its intricate dynamics spanning vast spatial and temporal scales, these simulations pose a complex and computationally challenging problem. Here, we propose an end‐to‐end deep learning model (RuptureNet2D) as a cost‐effective alternative (at evaluation time) to expensive numerical simulations. This model is trained on a data set of 300k simulations in two dimensions and is capable of simultaneously predicting two key earthquake source parameters: rupture time and final slip. Testing reveals exceptional model performance on faults with homogeneous and heterogeneous (with one or two asperities) frictional parameters but only requires a fraction (1/100,000 to 1/1,000) of the prediction time compared to numerical simulations. However, the accuracy of the neural network decreases as the fault length and number of asperities increases beyond the range of the training data. Nevertheless, our work demonstrates the applicability and efficiency of neural networks as a surrogate for earthquake dynamic rupture simulations which has a potential of significantly accelerating physics‐based earthquake source inversion and advancing our understanding of earthquake physics.

Application of convolutional neural networks for ultrasonic signal classification in the measurement of oil slick thickness on ocean surface

Application of convolutional neural networks for ultrasonic signal classification in the measurement of oil slick thickness on ocean surface

Classification of Samples via Neural-Network Augmented Two-Dimensional Infrared Spectroscopy.

The application of artificial neural network (ANN) techniques to spectroscopy has proven to be a powerful tool for the rapid and accurate classification of experimental samples. However, despite the unique abilities of two-dimensional infrared spectroscopy (2D IR), the use of ANNs to classify samples on the basis of their 2D-IR spectra has been unexplored. We present two investigations into utilizing ANNs to perform end-to-end classification of samples from their 2D-IR spectra. In the first, we construct a model that can perform a binary classification of experimental samples on the basis of their solvent. In the second, we demonstrate that classification is possible even for a single spectral slice of pump-delay and waiting-time combination even when samples display almost identical spectra. These results clearly demonstrate the potential of ANN-augmented 2D IR, with particular emphasis on its use as a technology for high-throughput screening applications.