School of Mechanical Engineering Seminar
Monday, December 31, 2018 at 14:00
Wolfson Building of Mechanical Engineering, Room 206

Efficient and robust forward simulation of complex geothermal processes

Denis Voskov

Department of Geoscience and Engineering, TU Delft

Department of Energy Resources Engineering, Stanford University

In the recent years, geothermal technology has received substantial attention as an alternative source of energy. However, the lack of detailed information about subsurface formations of interest often introduces significant uncertainties to the technological and economic planning of geothermal projects. That makes the robustness and efficiency of forward simulation extremely important. Geothermal modeling implies the solution of governing laws describing mass and energy transfer in the subsurface, which in turn requires the linearization of strongly nonlinear systems of equations. In my talk, I will describe a novel linearization strategy – Operator-Based Linearization (OBL) - capable to deal with complex nonlinear problems. The key idea of the approach is a transformation of discretised mass- and energy-conservation equations to an operator form with separate space-dependent and state-dependent operators. This transformation provides an opportunity for an approximate representation of the exact physics which is conceptually similar to an approximate representation of space and time discretization performed in conventional simulation. The current version of our simulation framework for geothermal applications includes OBL approximation for thermal-compositional physics of convection, conduction and buoyancy forces with multi-segmented wells and advanced nonlinear solvers.

The OBL approach enhances the computational performance of simulation and provides an opportunity for a simplified porting of simulation engine to heterogeneous computing architectures (such as GPU) which improves the performance even farther. I will show a few examples when the OBL framework is complemented with unique advanced simulation technologies such as compositional multiscale transport or species-based formulation for reactive transport. The advanced nonlinear solvers in OBL formulation is implemented based on the direct analysis of conventional operators in parameter space of nonlinear problem. In addition, the OBL approach can be used in creating physics-based data-driven models for complex subsurface processes. I will discuss several practical applications, ongoing developments and possible extensions of the proposed methodology.

Bio: Dr. Voskov is an Associate Professor at TU Delft. He received his PhD degree in Applied Mathematics from Gubkin’s Russian University of Oil and Gas in 2002. Dr. Voskov former positions include: senior researcher at Stanford University, founder and chief technology officer of Rock Flow Dynamics company, chief engineer at YUKOS EP company, and leading engineer-mathematician at the Institute for Problems in Mechanics within the Russian Academy of Sciences.

Speaker: Ben Fishman

M.Sc. student under the supervision of Prof. Hagit Messer-Yaron and Dr. Irit Opher

Wednesday, November 28^th, 2018 at 15:00

Room 011, Kitot Bldg., Faculty of Engineering

Prosodic Feature Criterion

Abstract

Prosody can be defined as the non-contextual information conveyed in speech utterances. Prosodic cues provide valuable information for human communication as prosody is used to express emotional states, attitudes and intentions. It carries clues regarding dialogue turns, phrase emphasis and even the physiological state of the speaker. Prosody has been researched extensively by linguists and speech scientists; However, little attention has been given to formulating and ranking the acoustic features that represent prosodic information.

In this work we present the Prosodic Feature Criterion (PFC) for evaluating the prosodic nature of a feature that was extracted from a speech signal. Using the PFC score we can rank the features, compare them and determine whether an acoustic or spectral feature carries prosodic information.

We explore the PFC using many kinds of features including ~4,300 features out of the OpenSMILE toolkit, which is a standard set of features widely used for acoustic analysis and prosody research. We use a few datasets in different languages; the main dataset is in Hebrew and was especially designed for research prosodic features. We apply our methodology successfully and find that prosodic features indeed are independent of the content of the utterance, while depend on prosodic manifestations.

We validate our methodology using several methods: showing that our ranking of prosodic features yields similar results to classification-based feature selection, showing visualization of the PFC idea using dimension reduction of multiple features representation, and comparing the PFC results to standard features, some of them are considered to be prosodic and some are not.

(The talk will be given in English)

Speaker:     Dr. Roy Schwartz
                   University of Washington and the Allen Institute for AI.

Monday, December 10^th, 2018
15:00 - 16:00

Room 011, Kitot Bldg., Faculty of Engineering

Towards Interpretable Deep Learning for Natural Language Processing

Abstract

Despite their superb empirical performance, deep learning models for natural language processing (NLP) are often considered black boxes, as relatively little is known as to what accounts for their success. This lack of understanding turns model development into a slow and expensive trial-and-error process, which limits many researchers from developing state-of-the-art models. Customers of deep learning also suffer from this lack of understanding, because they are using tools that they cannot interpret. In this talk I will show that many deep learning models are much more understandable than originally thought.

I will present links between several deep learning models and classical NLP models: weighted finite-state automata. As the theory behind the latter is well studied, these findings allow for the development of more interpretable and better-performing NLP models. As a case study, I will focus on convolutional neural networks (ConvNets), one of the most widely used deep models in NLP. I will show that ConvNets are mathematically equivalent to a simple, linear chain weighted finite-state automaton. By uncovering this link, I will present an extension of ConvNets that is both more robust and more interpretable than the original model. I will then present similar observations regarding six recently introduced recurrent neural network (RNN) models, demonstrating the empirical benefits of these findings to the performance of NLP systems.

This is joint work with Hao Peng, Sam Thomson and Noah A. Smith
 

Short Bio

Roy Schwartz is a postdoctoral researcher at the University of Washington and the Allen institute for AI. Roy's research focuses on improving deep learning models for natural language processing by gaining mathematical and linguistic understanding of these models. He received his Ph.D. and M.Sc. in Computer Science and his B.Sc. in Computer Science and Cognitive Science from the Hebrew University. Roy has won a best paper award at RepL4NLP 2018, as well as a Hoffman leadership and responsibility fellowship.