Search for: All records

Freak waves, waves significantly higher than neighboring waves, are a serious threat to ships and marine infrastructure. Despite significant refinement of operational wave models and recent progress in studying the theoretical foundations of such extreme events, the emergence of these events remains unpredictable. In this work, the authors propose a data-driven wave forecasting approach by combining the essence of common wave models, rapid oscillations, and slowly changing spectrum with data-driven techniques such as recurrent neural networks. A judicious minimization procedure is developed, wherein the sea surface elevation is first decomposed into harmonic functions with varying amplitudes. Then, the amplitude variations are forecasted by fitting universal, black-box models. This approach, which can be used to forecast wave crests and troughs in real time, is tested on available buoy data. Overall, the developed models and fitting strategies outperform simple benchmarks indicating the approach’s potential for operational, real-time wave forecasting. 

Generation and propagation of waves in a numerical wave tank constructed using Weakly Compressible Smoothed Particle Hydrodynamics (WCSPH) are considered here. Numerical wave tank simulations have been carried out with implementations of different Wendland kernels in conjunction with different numerical dissipation schemes. The simulations were accelerated by using General Process Graphics Processing Unit (GPGPU) computing to utilize the massively parallel nature of the simulations and thus improve process efficiency. Numerical experiments with short domains have been carried out to validate the dissipation schemes used. The wave tank experiments consist of piston-type wavemakers and appropriate passive absorption arrangements to facilitate comparisons with theoretical predictions. The comparative performance of the different numerical wave tank experiments was carried out on the basis of the hydrostatic pressure and wave surface elevations. The effect of numerical dissipation with the different kernel functions was also studied on the basis of energy analysis. Finally, the observations and results were used to arrive at the best possible numerical set up for simulation of waves at medium and long distances of propagation, which can play a significant role in the study of extreme waves and energy localizations observed in oceans through such numerical wave tank simulations. 

In the present effort, a data-driven modeling approach is undertaken to forecast aperiodic responses of non-autonomous systems. As a representative non-autonomous system, a harmonically forced Duffing oscillator is considered. Along with it, an experimental prototype of a Duffing oscillator is studied. Data corresponding to chaotic motions are obtained through simulations of forced oscillators with hardening and softening characteristics and experiments with a bistable oscillator. Portions of these datasets are used to train a neural machine and make response predictions and forecasts for motions on the corresponding attractors. The neural machine is constructed by using a deep recurrent neural network architecture. The experiments conducted with the different numerical and experimental chaotic time-series data confirm the effectiveness of the constructed neural network for the forecasting of non-autonomous system responses. 

Saliency methods are used extensively to highlight the importance of input features in model predictions. These methods are mostly used in vision and language tasks, and their applications to time series data is relatively unexplored. In this paper, we set out to extensively compare the performance of various saliency-based interpretability methods across diverse neural architectures, including Recurrent Neural Network, Temporal Convolutional Networks, and Transformers in a new benchmark † of synthetic time series data. We propose and report multiple metrics to empirically evaluate the performance of saliency methods for detecting feature importance over time using both precision (i.e., whether identified features contain meaningful signals) and recall (i.e., the number of features with signal identified as important). Through several experiments, we show that (i) in general, network architectures and saliency methods fail to reliably and accurately identify feature importance over time in time series data, (ii) this failure is mainly due to the conflation of time and feature domains, and (iii) the quality of saliency maps can be improved substantially by using our proposed two-step temporal saliency rescaling (TSR) approach that first calculates the importance of each time step before calculating the importance of each feature at a time step.