We integrate a multi-walled carbon nanotube (MWCNT)-based back contact to enhance hole extraction and device stability in perovskite solar cells (PSCs) without a hole transport layer (HTL), providing a cost-effective and stable alternative to conventional architectures by relying solely on an absorber layer and an electron transport layer (ETL). We propose a machine learning (ML)-driven framework that combines numerical simulations with experimental validation to simultaneously optimize efficiency and stability in HTL-free PSCs. We observe excellent agreement between a fabricated device and its simulated counterpart for some specific molar fraction. This work led to this Elsevier Sustainable Energy Technologies and Assessments publication.
In this work, we propose an end-to-end privacy-preserving CSI-based HAR framework that integrates a Convolutional Neural Network (CNN) with a temporal attention mechanism. We perform extensive evaluations on multiple benchmark datasets consisting of varying distance and height factors, as well as different environmental conditions. Our baseline non-privacy-preserving CNN–Temporal attention model achieves state-of-the-art performance. Additionally, we incorporate differential privacy (DP) into the training pipeline — enabling rigorous privacy guarantees through controlled noise injection and gradient clipping. We evaluate the proposed framework’s privacy–utility trade-off and demonstrate that even a strong privacy protection can maintain excellent recognition accuracy. Our framework can progressively approach the non-privacy-preserving performance for some parameter regime. This work led to this RSC Digital Discovery publication.
In this work, we present a comparative analysis of five different machine learning algorithms, namely Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Random Forest (RF), Naive Bayes (NB), and anomaly detection, in classifying different types of single-stage and multiple-stage PQDs. We generated a dataset consisting of 29 types of PQDs, while taking into consideration the variation of fundamental frequency. We note that variation of fundamental frequency can significantly affect the classification accuracy of the machine learning models. We introduced different levels of noise to the PQD data to mimic real-life scenarios. We used the Linear Discriminant Analysis (LDA) for dimensionality reduction and feature extraction. Our results show that the classification accuracy decreases as the noise level is increased and that a decrease in the classification accuracy is possible when the variation of the fundamental frequency is considered. This work led to this Measurement publication.
Photovoltaic (PV) panels can develop many physical faults—like cracks, dirt, shading, or temperature damage—that reduce their power output. Detecting these faults early is important for keeping solar systems efficient. In this work, we present a machine learning method that detects PV faults using unlabeled electroluminescence (EL) images. We first label the images automatically using k‑means clustering, with features extracted from a pre‑trained VGG‑16 model. Principal component analysis (PCA) shows that 64 components capture most of the important information. This led to this Energy Advances publication.
Detecting deepfakes is increasingly important, but it is difficult because deepfakes can be created in many different ways. In the 2025 IEEE Signal Processing Cup, the dataset included many types of deepfake images but had a large imbalance between real and fake samples. To address this, we developed a multistage training approach using CAE‑Net, an ensemble model that combines three architectures: EfficientNet, a Data‑Efficient Image Transformer, and ConvNeXt with wavelet transforms. Our method achieved 94.63% accuracy and 97.37% AUC on the competition’s validation set. This work led to this Journal of Visual Communication and Image Representation publication.
As data grows in areas like healthcare, finance, and security, it offers many benefits but also raises serious privacy concerns because personal identity can often be inferred even from anonymized data. Existing machine‑learning and data‑sharing methods struggle to protect privacy when dealing with high‑dimensional data. To address this, we propose a new data‑publishing method called DP‑CDA. It creates synthetic data by carefully mixing sensitive data within each class and adding controlled randomness to guarantee strong privacy. Our analysis shows that DP‑CDA offers better privacy protection than existing methods while keeping the data more useful. This work led to this Security and Privacy publication.
Recommendation systems use a lot of user data, which can risk exposing sensitive personal information. Differential privacy can protect users, but when applied to neural‑network‑based recommenders, it often reduces accuracy—creating a trade‑off between privacy and performance. Traditional matrix‑factorization ethods struggle to balance this trade‑off. In this work, we propose a neural‑network‑based collaborative filtering model that achieves both strong privacy and high accuracy. Using Rényi differential privacy, we measure how much privacy is lost during training and test the model on real datasets under different privacy settings. Our results show that the proposed ANN model provides strong privacy guarantees while maintaining excellent recommendation quality, outperforming many existing private and non‑private methods. This work led to this Measurement publication.
We deveeloped fast machine‑learning method to predict the full electronic band structure of monolayer TMD alloys, which are usually computed using computationally expensive DFT simulations. Using DFT data for alloys made from W, Mo, and S/Se/Te, the researchers trained an “extra trees” model to predict both conduction and valence bands. The model identifies important factors—like the ratio of different atoms—that shape the band structure. It performs very accurately and even works well for new compositions not included in the training set. Unlike models that only predict band gaps, this approach provides complete band information, enabling the calculation of additional useful properties such as effective carrier mass. Overall, the model offers a fast and informative alternative to full DFT simulations for materials research. This work led to this Materials Advances publication.
Developed a novel function computation scheme Gaussian Functional mechanism that ensures strict privacy guaratnee and offers much better utility than existing methods. Additionally, we extend our novel mechanism such that it works seemlessly in decentralized-data settings and propose capeFM, which offers the same privacy-utility trade-off as centralized schemes. This theoretical work is also validated on several real datasets for regression and classification. This work led to this Entropy Information Theory, Probability and Statistics publication.
Developed an efficient, novel, and implementation-friendly framework for human activity recognition (HAR) using Channel State Information (CSI) of WiFi signal. The framework employs wavelet-based feature extraction and principal component analysis (PCA) based subcarrier fusion to achieve excellent recognition performance of human activity on multiple real datasets, that include variety of environmental conditions. This work led to this Elsevier Digital Signal Processing publication.