Fadi Al Machot

Platform motion in floating offshore wind turbines alters wake recovery and coherent structures. High-fidelity CFD can resolve these effects, but repeated simulations remain computationally expensive. In direct full-field surrogate prediction, motion-induced wake changes are learned together with the dominant fixed-wake background, making them less explicit. Here, wake prediction is formulated as a perturbation-learning problem by decomposing the moving-turbine wake into a time-resolved fixed-turbine reference wake and a motion-induced perturbation. A motion-conditioned recurrent neural network predicts this perturbation from fixed-wake and perturbation histories together with motion type, amplitude, Strouhal number, and phase. The moving wake is reconstructed by adding the predicted perturbation to the fixed-reference wake. The method is evaluated using high-fidelity CFD simulations of a model floating wind turbine undergoing prescribed pitch and roll motions. Compared with direct moving-wake prediction, the perturbation formulation reduces normalized velocity-deficit RMSE by 49.1% for representative cases and wake-center error by 43.6%. Because the same fixed-reference wake is used in reconstruction, the perturbation-error reduction carries to the reconstructed wake. The predicted fields identify where platform motion modifies wake deficit, shear-layer structure, recovery, and displacement. Lower errors are also obtained for velocity profiles, maximum deficit, and wake width, supporting a more accurate and interpretable wake-reconstruction approach.

Accurate and stable prediction of unsteady flows around moving boundaries remains challenging for data-driven models because of strong temporal dependencies, parametric forcing, and error accumulation during long-horizon integration. Existing approaches often treat moving-boundary flow prediction as snapshot-to-snapshot mapping or time-dependent regression, without explicitly introducing a dynamical state driven by prescribed body motion. This work proposes a motion-driven latent dynamical system formulation in which prescribed motion acts as a time-dependent external forcing of the latent flow evolution. The proposed Motion-Conditioned Physics-integrated Recurrent Neural Network (MoC-PiRNN) uses an encoder–decoder architecture to construct a spatial latent state, while a recurrent transition mapping advances this state under motion conditioning. A latent mixing block is included to support global information exchange beyond local convolutional recurrence, and physics-integrated regularization is used to improve physical consistency without deriving an explicit reduced-order model. From an artificial intelligence perspective, the contribution is a motion-conditioned recurrent learning framework that embeds exogenous forcing directly into latent-state evolution. From an engineering perspective, the method provides a surrogate model for stable long-horizon prediction of unsteady moving-boundary flows. The framework is evaluated on a two-dimensional heaving square cylinder simulated using overset-grid computational fluid dynamics. Numerical results show improved long-horizon stability across sampled motion amplitudes and frequencies. For the present case, the physics-integrated formulation reduces flow-field prediction error by more than 20% and lift-coefficient root-mean-square error by more than 40% relative to the motion-only model, demonstrating reduced error accumulation during autoregressive rollouts.

Accurate and stable prediction of unsteady flows around moving boundaries remains challenging for data-driven models due to strong temporal dependencies, parametric forcing, and error accumulation during long-horizon integration. Existing approaches typically treat moving-boundary flow prediction as either a sequence of snapshot-to-snapshot mappings or as a time-dependent regression problem. While these methods have shown success for short-term prediction, they typically do not introduce an explicit dynamical state whose evolution is directly driven by prescribed body motion. In this work, we propose a motion-driven latent dynamical system formulation for moving-boundary flows, in which prescribed body motion serves as a time-dependent external forcing that drives the evolution of the latent flow state. The proposed Motion-Conditioned Physics-integrated Recurrent Neural Network (MoC-PiRNN) introduces a spatial latent state via an encoder-decoder architecture, with temporal evolution represented by a learned recurrent-state transition mapping. Prescribed body motion is incorporated as a time-dependent exogenous forcing signal that explicitly drives the latent-state evolution at each recurrent step. In addition, a latent mixing block is introduced to enable global information exchange beyond purely convolutional recurrent dynamics. Physics-integrated regularization is incorporated to produce physically consistent predictions without requiring an explicit derivation of the reduced-order model. The method is evaluated on a canonical moving-boundary problem involving a two-dimensional heaving square cylinder simulated using overset-grid CFD. Numerical experiments demonstrate improved long-horizon stability of the predicted flow evolution, while maintaining competitive performance across a sampled range of motion amplitudes and frequencies.In particular, for the present case, the physics-integrated formulation reduces the flow-field prediction error by more than 20% and improves the accuracy of lift coefficient predictions by more than 40%. The results indicate that the proposed formulation is effective in mitigating error accumulation during temporal rollouts and highlight the remaining challenges in generalizing to unseen motion conditions.

Universal deepfake detection aims to identify AI-generated images across a broad range of generative models, including unseen ones. This requires robust generalization to new and unseen deepfakes—which emerge frequently—while minimizing computational overhead to enable large-scale deepfake screening, a critical objective in the era of Green AI. In this work, we explore frequency-domain masking as a training strategy for deepfake detectors. Unlike traditional methods that rely heavily on spatial features or large-scale pretrained models, our approach introduces random masking and geometric transformations, with a focus on frequency masking due to its superior generalization properties. We demonstrate that frequency masking not only enhances detection accuracy across diverse generators but also maintains performance under significant model pruning, offering a scalable and resource-conscious solution. Our method achieves state-of-the-art generalization on GAN- and diffusion-generated image datasets and exhibits consistent robustness under structured pruning. These results highlight the potential of frequency-based masking as a practical step toward sustainable and generalizable deepfake detection. Code and models are available at https://github.com/chandlerbing65nm/FakeImageDetection .

Long-term time series forecasting often relies on frequency- or time-domain decomposition to capture trends and seasonalities. Frequency based methods, such as the Discrete Fourier Transform, isolate dominant periodic components and suppress noise, but they operate in the complex domain. Time-domain models apply trend–seasonality decomposition but often capture only a single seasonality entangled with residual noise. To address these limitations, we propose the \ac{FragFM}-Fragmented Fourier Matrix--a time domain framework that incorporates spectral insights through fixed Fourier fragments. FragFM removes trends via moving averages, extracts multiple seasonal components with fragmented Fourier bases, and corrects residual noise. Operating fully in the real domain, FragFM achieves forecasting accuracy comparable to leading frequency-domain models while consistently outperforming traditional time-domain approaches. Its lightweight, interpretable design offers an efficient solution for long-term forecasting, bridging time- and frequency-domain methods. The code repository is: https://github.com/K0DX-DV1NUX/FragFM

Early-stage diagnosis of paddy diseases—based on initial visible symptoms—is crucial for reducing chemical use and preventing disease spread. Automated paddy disease diagnosis facilitates this, and helps to improve crop production. Proper segmentation of diseased regions is crucial for precise severity analysis. Segmentation models can be used to address the challenges related to the complex disease patterns. This study analyses the performance and applicability of seven classification models (DenseNet, InceptionV3, Xception, MobileNet, Vision Transformer (ViT), parallel-ViT, and a custom ensemble model of DenseNet121 and Xception) and eight segmentation models (DeepLabv3+, UNet, VGG16 UNet, Inception-ResNetV2 UNet, TransUNet, Attention UNet, MultiResUNet, and Deep Residual UNet) for this domain. A comparative analysis of these models highlighting their structural differences, advantages, and limitations has also been provided. Given the high computational demands and high data dependency of ViTs, we evaluated whether traditional models, their ensembles, or memory-efficient models could match ViT performance and found it feasible. Additionally, we explore the effect of augmentation intensity on the classification models. We also identify significant research gaps within this domain. To address the lack of open-access paddy disease segmentation datasets, we created a novel paddy disease segmentation dataset using image-processing techniques. The applicability of the above-mentioned segmentation models has been evaluated for paddy diseases using this newly created dataset and we have identified that Deep Residual UNet is the most suitable model to be used in resource constraint applications considering its quantitative and qualitative performance, model size, and structural advantages and limitations. Furthermore, we have investigated the impact of training these segmentation models with a significantly higher number of augmented images more than double the original dataset. We observed that while the quantitative performance increased with increased data, the qualitative performance degraded in a few cases. Our results will help researchers in developing effective autonomous systems for paddy disease management.