Data Scientist @ Emulate Energy | Asst. Prof. in Computer Science @ University of Sarajevo
Polje Istraživanja: Numerical simulations Machine learning Mathematical optimization
In offshore engineering design, nonlinear wave models are often used to propagate stochastic waves from an input boundary to the location of an offshore structure. Each wave realization is typically characterized by a high-dimensional input time-series, and a reliable determination of the extreme events is associated with substantial computational effort. As the sea depth decreases, extreme events become more difficult to evaluate. We here construct a low-dimensional characterization of the candidate input time series to circumvent the search for extreme wave events in a high-dimensional input probability space. Each wave input is represented by a unique low-dimensional set of parameters for which standard surrogate approximations, such as Gaussian processes, can estimate the short-term exceedance probability efficiently and accurately. We demonstrate the advantages of the new approach with a simple shallow-water wave model based on the Korteweg–de Vries equation for which we can provide an accurate reference solution based on the simple Monte Carlo method. We furthermore apply the method to a fully nonlinear wave model for wave propagation over a sloping seabed. The results demonstrate that the Gaussian process can learn accurately the tail of the heavy-tailed distribution of the maximum wave crest elevation based on only $$1.7\%$$ 1.7 % of the required Monte Carlo evaluations.
We model shallow-water waves using a one-dimensional Korteweg–de Vries equation with the wave generation parameterized by random wave amplitudes for a predefined sea state. These wave amplitudes define the high-dimensional stochastic input vector for which we estimate the short-term wave crest exceedance probability at a reference point. For this high-dimensional and complex problem, most reliability methods fail, while Monte Carlo methods become impractical due to the slow convergence rate. Therefore, first within offshore applications, we employ the dimensionality reduction method called Active-Subspace Analysis. This method identifies a low-dimensional subspace of the input space that is most significant to the input–output variability. We exploit this to efficiently train a Gaussian process (i.e., a kriging model) that models the maximum 10-min crest elevation at the reference point, and to thereby efficiently estimate the short-term wave crest exceedance probability function. The active low-dimensional subspace for the Korteweg–de Vries model also exposes the expected incident wave groups associated with extreme waves and loads. Our results show the advantages and the effectiveness of the active-subspace analysis against the Monte Carlo implementation for offshore applications.
We demonstrate that a single 6mm line sample of simulated near-field speckle intensity suffices for accurate estimation of the concentration of dielectric micro-particles over a range from 104 to 6⋅106 particles per ml. For this estimation, we analyze the speckle using both standard methods (linear principal component analysis, support vector machine (SVM)) and a neural network, in the form of a sparse stacked autoencoder (SSAE) with a softmax classifier or with an SVM. Using an SSAE with SVM, we classify line speckle samples according to particle concentration with an average accuracy of over 78%, with other methods close behind.
Ova stranica koristi kolačiće da bi vam pružila najbolje iskustvo
Saznaj više