Mohan Kumar
Gajendran

Abstract

Cerebrospinal fluid (CSF) dynamics within the optic nerve subarachnoid space (ONSAS) are increasingly recognised as potential factors contributing to the pathogenesis of optic neuropathies, including normal-tension glaucoma (NTG) and idiopathic intracranial hypertension (IIH). In this study, high-resolution T2-weighted magnetic resonance imaging (MRI) data were manually segmented and processed for greyscale analysis and pixel-intensity mapping, enabling anatomically accurate three-dimensional reconstructions of the ONSAS. These reconstructions were imported into OpenFOAM software for computational fluid dynamics (CFD) simulations, incorporating physiologically realistic inlet pressure waveforms, boundary conditions, tissue porosity, and lymphatic drainage characteristics in the near optic nerve head portion of ONSAS, called the distal portion. Solver selection in OpenFOAM was optimised for the narrow geometry and unique boundary features of the ONSAS, yielding high-resolution maps of CSF flow dynamics. Multiple drainage scenarios were simulated to represent healthy, NTG, and IIH conditions. Area-averaged CSF velocities were quantified for each condition and compared with the scarce and available values reported in phase-contrast MRI (PC-MRI) studies. The results provide novel, physiologically relevant insights into CSF transport behaviour in the ONSAS under different physiological and pathophysiological conditions, with potential implications for the diagnosis, management, and treatment of optic neuropathies.

Abstract

This paper presents a systematic evaluation of lexical, dense, and hybrid retrieval pipelines for retrieval-augmented generation (RAG) under matched, reproducible conditions. Using the BEIR Natural Questions benchmark with judged-document-aware sampling, we jointly assess retrieval effectiveness (nDCG, recall, MRR) and system efficiency (runtime and memory). Results show that dense sentence-transformer encoders substantially outperform BM25 in ranking quality (nDCG@10=0.916 vs. 0.799) and early precision, while approximate nearest-neighbor search (HNSW) preserves these gains with negligible overhead. Hybrid reciprocal-rank fusion increases coverage (Recall@100=0.997) but offers limited gains in ranking accuracy, and pseudo-relevance feedback (BM25+RM3) under-performs the lexical baseline. Cross-encoder reranking achieves the highest overall effectiveness (nDCG@10=0.954, MRR@10=0.948) at more than 50× the runtime of dense retrieval. Statistical tests confirm all pairwise differences among BM25, RM3, and dense retrievers as significant (p<0.01). The study provides quantitative guidance for balancing retrieval quality and cost in RAG pipelines, showing that dense-hybrid retrieval offers strong accuracy-efficiency trade-offs for practical deployments.

Abstract

The predictive value of deep-learning models in finance depends critically on the temporal context supplied to the network, yet the literature lacks a systematic, multi-market evaluation of how input-sequence length alone shapes forecast accuracy. This study investigates the effect of ten look-back windows (1 to 252 trading days) on next-day returns for eight major equity indices spanning North America, Europe, and Asia-Pacific. A uniform experimental grid of 10 × 2 × 8 combinations of look-back horizon, architecture (bidirectional LSTM and GRU), and index was trained on daily OHLCV data from April 2017 to March 2025. Results reveal that optimal window length is market-specific: momentum-dominated indices (e.g., S&P 500, TSX Composite) reach minimum RMSE with 1 to 3-day histories, whereas macro-trend markets (Nikkei 225, Shanghai Composite, SENSEX) require six-month to annual windows for peak performance. GRUs marginally outperform LSTMs at ultra-short horizons, but LSTMs regain the lead once the window exceeds one month. These findings challenge the common short-is-best heuristic and provide actionable guidance for practitioners: tune look-back length per market, pair GRUs with sub-week windows for efficiency, and reserve LSTMs for macro-driven horizons.

Abstract

Abstract

The invention uses electroretinography (ERG) signals in a machine-learning-based framework to detect early-stage glaucoma. The system extracts statistical and wavelet-based features from ERG signals, then trains machine learning models to provide binary classification (glaucomatous vs. non-glaucomatous), distinguish glaucoma progression stages through multiclass classification, and predict retinal ganglion cell (RGC) counts for visual functionality assessment. The framework improves upon existing diagnostic methods by detecting functional deficits of early and various stages of glaucoma with high accuracy through preprocessing, feature extraction, and model training techniques.

Abstract

Selecting the input look-back window is a critical yet often under-examined design choice in short-horizon foreign exchange forecasting. This paper quantifies how look-back length affects one-day-ahead performance of gated recurrent networks across eight widely traded USD currency pairs from April 2017 to March 2025. Using only OHLC prices and log returns, we run a controlled grid of 160 experiments that crosses ten window lengths (1–252 trading days) with two architectures (LSTM and GRU) under a uniform rolling evaluation protocol. Three findings emerge. First, short windows dominate: windows of 1, 3, or 7 days yield the lowest MAPE in most configurations. Second, the optimal window is pair-specific: moderately longer context can help for smoother or more policy-managed pairs (for example, USD/CNY), while freely floating pairs (for example, USD/JPY) favor the shortest context. Third, longer windows (90–252 days) rarely help and often degrade accuracy. The results recommend treating look-back length as a first-class, per-instrument hyperparameter and caution against assuming that more historical context improves next-day forecasts.

Abstract

Renewable energy sources are essential to address climate change, fossil fuel depletion, and stringent environmental regulations. Horizontal-axis wind turbines (HAWTs) are particularly suited to meet this demand, but their efficiency is affected by adverse weather conditions like rain. This study investigates wind turbine power prediction under rainy conditions by integrating Blade Element Momentum (BEM) theory with explainable artificial intelligence (XAI). The S809 airfoil's aerodynamic characteristics, used in NREL wind turbines, were analyzed using ANSYS FLUENT and symbolic regression under varying rain intensities. Simulations at a Reynolds number of 1×10⁶ were performed using the Discrete Phase Model (DPM) and k–ω SST turbulence model, with liquid water content (LWC) values of 0 (dry), 10, 25, and 39 g/m³. The results indicated that rain led to reduced lift and increased drag. Machine learning models predict changes in the airfoil coefficients under rain with an R² value of 0.97. The proposed XAI framework models rain effects at a lower computational time, enabling efficient wind farm performance assessment. It was found that a heavy rain LWC of 39 g/m³ could reduce power output by 5.7% to 7%.

Abstract

The morphological abnormalities in the retinal blood vessel have a close association with cerebrovascular, cardiovascular, and systemic diseases. It makes the retinal artery/vein (A/V) classification salient for clinical decision-making. The existing methods find it challenging to correctly classify A/V with non-uniform brightness and vessel thickness, especially at the bifurcation and endpoints. To avoid these problems and increase precision, AV-Net is proposed. It uses the context information and performs data fusion to improve A/V classification. Specifically, the AV-Net offers a module that fuses local and global vessel information for creating a weight map to constrain the A/V features. It helps suppress the background-prone features and improve region extraction at the bifurcation and endpoints. In addition, to improve model robustness, the AV-Net uses a multiscale-feature module that captures coarse and fine details.

Abstract

This chapter provides a detailed analysis of the operation of mixed-mode ventilation during epidemics, concentrating on the pivotal role of indoor air quality (IAQ). It underlines the importance of ventilation in IAQ management, particularly for airborne infection control. However, our principal focus is mixed-mode ventilation, a combined approach of natural and mechanical methods, which we highlight as promising for IAQ management, airborne disease control, and also energy-saving solutions. Our examination includes multiple case studies for each diverse environment, such as educational buildings, hospitals, office buildings, and residential buildings, each evaluated through different methods, including computational fluid dynamics and experimental approaches. Our observations illustrate the significant role of efficient ventilation in improving IAQ, mitigating airborne infection risks, and enhancing occupant comfort, especially during epidemics.

Abstract

As wind energy continues to be a crucial part of sustainable power generation, the need for precise and efficient modeling of wind turbines, especially under yawed conditions, becomes increasingly significant. This study introduces a machine learning-based symbolic regression approach for elucidating wake dynamics. Utilizing WindSE's actuator line method (ALM) and Large Eddy Simulation (LES), we model an NREL 5-MW wind turbine under yaw conditions ranging from no yaw to 40 degrees. Leveraging a hold-out validation strategy, the model achieves robust hyper-parameter optimization, resulting in high predictive accuracy in predicting wake deflection and velocity deficit at both the wake center and hub height. Our approach provides an accurate and computationally efficient tool for wind farm optimization, filling a significant gap in the literature on machine-learning-based wake models for yaw wake prediction.

Abstract

This chapter provides a comprehensive exploration of the evolving role of machine learning in Indoor Environmental Quality (IEQ) assessment. As urban living spaces become increasingly enclosed, the importance of maintaining optimal IEQ for human health and well-being has surged. Traditional methods for IEQ assessment, while effective, often fail to provide real-time monitoring and control. This gap is increasingly being addressed by the integration of machine learning techniques, allowing for enhanced predictive modeling, real-time optimization, and robust anomaly detection. The chapter delves into a comparative analysis of various machine learning techniques including supervised, unsupervised, and reinforcement learning, demonstrating their unique benefits in IEQ assessment. Practical implementations of these techniques in residential, commercial, and specialized environments are further illustrated through detailed case studies.

Abstract

LNG stored at −160 °C holds a significant amount of exergy, which is released during regasification (LNG cold energy). Heat from ocean water and additional gas burners are commonly used for regasification, dumping available cold energy and creating an aquatic microclimate with detrimental environmental consequences on marine life. Conversely, excessive CO₂ emissions from flue gas substantially impact the environment. Cryogenic carbon capture (CCC) offers high CO₂ recovery rates and purity, but producing the required cryogenic cold energy is costly. The objective of this study is to efficiently exchange waste cold energy from LNG with waste heat energy from flue gas: the cryogenic cold energy from LNG is utilized to extract CO₂ from flue gas by liquefying it, while the heat energy from flue gas regasifies the LNG. CFD simulations of a shell-and-tube heat exchanger were conducted, accounting for the phase changes of both streams — a proof of concept contributing toward Net Zero Carbon Emissions by 2050.

Abstract

Wind turbines are often clustered together due to limited suitable locations, necessitating the study of wake aerodynamics to optimize wind farm performance. Understanding and predicting wake characteristics — decreased velocity deficit and increased turbulence intensity — is crucial for improving the efficiency of downstream turbines. Recent advancements in CFD have enabled more accurate predictions, albeit at the cost of increased computational time. Machine-learning-based wake models leverage CFD data for rapid predictions, but have been limited by their reliance on a single wind turbine's CFD results and their lack of generalizability to other turbines with different blade geometries and operating conditions. In response, this study presents a novel wake prediction model using the symbolic regression technique. This approach offers precise aerodynamic performance predictions while maintaining simplicity, interpretability, and generalizability across various wind turbine designs, facilitating better decision-making in turbine placement and operation.

Abstract

Abstract

Wind turbine blades experience excessive load due to inaccuracies in the prediction of aerodynamic loads by conventional methods during design. The blade element momentum (BEM) method is the best-known design tool for evaluating aerodynamic performance due to its simplicity and short processing time. As the turbine rotates, the aerofoil lift coefficient is enhanced — notably in the rotor's inboard section — relative to 2D predictions, induced by centrifugal pumping action and the Coriolis force, delaying the occurrence of stall. This rotational effect significantly influences aerodynamic performance but is not captured by BEM. Existing stall delay correction models either fail or inaccurately predict the lift enhancement. This paper proposes a novel stall delay correction model based on symbolic regression for precise aerodynamic performance prediction by the BEM process, generated from the disparity in 3D and 2D aerofoil coefficients for the NREL Phase VI turbine and validated against experimental results of the MEXICO rotor.

Abstract

Machine Learning (ML) algorithms have become more prevalent in recent years, and they are being used to tackle complicated issues across a broad range of fields. Wind energy is not an exception, as ML has recently been applied to wind turbine blade design, wake velocity and wake turbulence intensity prediction, and even wind farm optimization. The immense learning ability of ML models enables them to be trained to predict and regress a complex relationship with a high degree of accuracy. However, data for testing ML models often originate from the same rotor simulation used for training, with just slight variations in operating conditions. This research investigates the generalizability of ML-based wake prediction models — whether ML can correctly predict wake properties using data from a different wind turbine that was not used for training. The observations show that a generalized ML wake model requires training data from multiple turbines with a wide range of operating conditions. In addition, advanced regularization, complex loss functions, and ML methods that focus on capturing the physics (such as Physics-Informed Neural Networks and symbolic regression) can be utilized.

Abstract

Deep learning has an enormous impact on medical image analysis. Many computer-aided diagnostic systems equipped with deep networks are rapidly reducing human intervention in healthcare. Among several applications, medical image semantic segmentation is one of the core areas of active research to delineate anatomical structures and other regions of interest, with significant contributions to guided interventions, radiotherapy, and improved radiological diagnostics. This article provides an overview of deep convolutional neural architecture, platforms and applications of deep neural networks, metrics used for empirical evaluation, state-of-the-art semantic segmentation architectures based on foundational convolution concepts, and a review of publicly available medical image datasets highlighting four distinct regions of interest, along with open-ended potential research directions.

Abstract

State-of-the-art deep neural networks trained on large amounts of data can diagnose cardiac arrhythmias better than cardiologists, but the requirement of high-volume training data is not pragmatic. In this research, the identification and classification of three ECG patterns are analyzed from a transfer learning perspective. Features learned from general image classification are transferred to time-series ECG signal classification: various modern deep networks trained on the ImageNet database are re-utilized for classifying scalograms (2D representations) of ECG signals, and their performance on ECG classification is assessed.

Abstract

A nozzle is used to convert pressure energy to kinetic energy in order to produce thrust. For subsonic flows a convergent nozzle is used, and for supersonic flows a convergent–divergent (CD) nozzle is employed, in which flow is accelerated from low subsonic to sonic velocity at the throat and further expanded to supersonic velocities at the exit. This paper focuses on designing a de Laval nozzle to attain supersonic flow and optimizing it to achieve maximum thrust without flow separation due to shock waves. For maximum thrust and efficiency the direction of the flow must be parallel to the axis of the nozzle. A CD nozzle is designed with the help of the G.V.R. Rao method, based on the Method of Characteristics, and analyzed using CFD for various performance parameters. Flow conditions were selected based on the pressure, temperature, and gases available at the exit of the combustion chamber. At the exit of the nozzle, shock-induced flow separation due to overexpansion and optimum expansion conditions are studied.

Abstract

The aerofoil is a primary element to be designed during the initial phase of creating any new aircraft — the component that forms the cross-section of the wing. It took about half a century to break the sound barrier after man started to fly; the difficulties were the infinite drag experienced in the transonic range and the inability to sustain lift at higher Mach numbers. A supersonic flight requires a wing capable of maintaining lift at high Mach numbers and inducing oblique and bow shocks at optimum Mach numbers. Critical speed — the velocity at which there is sudden loss in Mach number and lift — is a function of the aerofoil's shape, angle of attack, and thickness. In this paper, the supersonic aerofoil NACA 66-206 is designed and analyzed using CFD to study the shock and aerodynamic properties at Mach numbers 1.2, 1.5, 2.0, 2.5, and 3.0, with the objective of achieving minimal wave drag at high Mach numbers while sustaining lift.

Abstract

Abstract

Abstract

This study presents a novel educational framework developed to enhance the interpretative skills of graduate students in neuroscience. The framework integrates advanced artificial intelligence technologies, including Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG), aiming to improve students' engagement with and understanding of complex research articles. The framework utilizes OpenAI's LLMs for in-depth semantic analysis and synthesis of article content, supported by the Mathpix API for transforming unstructured PDF documents into analyzable formats. RAG, anchored in a key neuroscience textbook, enhances the contextual accuracy of the LLM outputs. Additionally, OpenAI's vision capabilities are employed to analyze visual data within articles, such as images and plots. The learning experience is customized through a diagnostic interface that allows students to adjust their knowledge level. The use of RAG led to noticeably improved comprehension of text-based content, significantly enhancing student engagement and understanding. These findings indicate the framework's strong potential in academic settings and demonstrate the powerful role of AI in educational innovation.

Abstract

Abstract

In this study, we performed a comparative analysis of machine learning models for early-stage glaucoma detection. The comparison examines models using engineered and extracted statistical and wavelet-based features against models that also incorporate medically relevant features from electroretinograms (ERGs) — A-wave amplitude, B-wave amplitude, A/B ratio, and the implicit times of the A and B waves. The dataset included ERG signals from eyes of DBA/2 mice (15 four-month-old and 15 eleven-month-old males), a model of pigmentary glaucomatous optic neuropathy characterized by chronic retinal neurodegeneration that increases in severity with age. Signals were grouped based on age and intraocular pressure, and the extracted features were used to train various machine learning models. Models integrating medically relevant ERG features significantly improved in performance — the ensemble bagged method displayed superior effectiveness in classifying ERG signals in both binary and multiclass setups. Our study demonstrates that the inclusion of medically relevant ERG features greatly enhances performance in early-stage glaucoma detection, with significant potential for identifying functional deficits across varying stages of glaucoma in mouse models and potentially also in human patients.