Beginning immediately, until at least April 30, all seminars will take place virtually, through Zoom. Links to connect can be found in the schedule below once they are created.
In the Spring 2019 Semester, the Center of Mathematical Sciences and Applications will be hosting a seminar on Fluid Dynamics. The seminar will take place on Wednesdays from 3:00-4:00pm in CMSA G10.
|2/25/2020||Keaton Burns, MIT||Title: Flexible spectral simulations of low-Mach-number astrophysical fluids
Abstract: Fluid dynamical processes are key to understanding the formation and evolution of stars and planets. While the astrophysical community has made exceptional progress in simulating highly compressible flows, models of low-Mach-number stellar and planetary flows typically use simplified equations based on numerical techniques for incompressible fluids. In this talk, we will discuss improved numerical models of three low-Mach-number astrophysical phenomena: tidal instabilities in binary neutron stars, waves and convection in massive stars, and ice-ocean interactions in icy moons. We will cover the basic physics of these systems and how ongoing additions to the open-source Dedalus Project are enabling their efficient simulation in spherical domains with spectral accuracy, implicit timestepping, phase-field methods, and complex equations of state.
|9/18/2019||Jiawei Zhuang (Harvard)||Title: Simulation of 2-D turbulent advection at extreme accuracy with machine learning and differentiable programming
Abstract: The computational cost of fluid simulations grows rapidly with grid resolution. With the recent slow-down of Moore’s Law, it can take many decades for 10x higher resolution grids to become affordable. To break this major barrier in high-performance scientific computing, we used a data-driven approach to learn an optimal numerical solver that can retain high-accuracy at much coarser grids. We applied this method to 2-D turbulent advection and achieved 4x effective resolution than traditional high-order flux-limited advection solvers. The machine learning component is tightly integrated with traditional finite-volume schemes and can be trained via an end-to-end differentiable programming framework. The model can achieve near-peak FLOPs on CPUs and accelerators via convolutional filters.
|9/25/2019||Yantao Yang (Peking University)||Title: Double diffusive convection and thermohaline staircases
Abstract: Double diffusive convection (DDC), i.e. the buoyancy-driven flow with fluid density depending on two scalar components, is omnipresent in many natural and engineering environments. In ocean this is especially true since the seawater density is mainly determined by temperature and salinity. In upper water of both (sub-) tropical and polar oceans, DDC causes the intriguing thermohaline staircases, which consist of alternatively stacked convection layers and sharp interfaces with high gradients of temperature and salinity. In this talk, we will focus on the fingering DDC usually found in (sub-)tropical ocean, where the mean temperature and salinity decrease with depth. We numerically investigate the formation and the transport properties of finger structures and thermohaline staircases. Moreover, we show that multiple states exit for the exactly same global condition, and individual finger layers and finger layers within staircases exhibit very different transport behaviors.
|10/9/2019||Samuel Rudy (MIT)||Title: Data-driven methods for discovery of partial differential equations and forecasting
Abstract: A critical challenge in many modern scientific disciplines is deriving governing equations and forecasting models from data where derivation from first principals is intractable. The problem of learning dynamics from data is complicated when data is corrupted by noise, when only partial or indirect knowledge of the state is available, when dynamics exhibit parametric dependencies, or when only small volumes of data are available. In this talk I will discuss several methods for constructing models of dynamical systems from data including sparse identification for partial differential equations with or without parametric dependencies and approximation of dynamical systems governing equations using neural networks. Limitations of each approach and future research directions will also be discussed.
|10/23/2019||Kimee Moore (Harvard)||Title: Using magnetic fields to investigate Jupiter’s fluid interior
Abstract: The present-day interior structure of a planet is an important reflection of the formation and subsequent thermal evolution of that planet. However, despite decades of spacecraft missions to a variety of target bodies, the interiors of most planets in our Solar System remain poorly constrained. In this talk, I will discuss how actively generated planetary magnetic fields (dynamos) can provide important insights into the interior properties and evolution of fluid planets. Using Jupiter as a case study, I will present new results from the analysis of in situ spacecraft magnetometer data from the NASA Juno Mission (currently in orbit about Jupiter). The spatial morphology of Jupiter’s magnetic field shows surprising hemispheric asymmetry, which may be linked to the dissolution of Jupiter’s rocky core in liquid metallic hydrogen. I also report the first definitive detection of time-variation (secular variation) in a planetary dynamo beyond Earth. This time-variation can be explained by the advection of Jupiter’s magnetic field by the zonal winds, which places a lower bound on the velocity of Jupiter’s winds at depth. These results provide an important complement to other analysis techniques, as gravitational measurements are currently unable to uniquely distinguish between deep and shallow wind scenarios, and between solid and dilute core scenarios. Future analysis will continue to resolve Jupiter’s interior, providing broader insight into the physics of giant planets, with implications for the formation of our Solar System.
|11/6/2019||Federico Toschi (Eindhoven University of Technology)||Title: Deep learning and reinforcement learning for turbulence
Abstract: This talk tells two stories.
Chapter 1: We investigate the capability of a state-of-the-art deep neural model at learning features of turbulent velocity signals. Deep neural network (DNN) models are at the center of the present machine learning revolution. The set of complex tasks in which they over perform human capabilities and best algorithmic solutions grows at an impressive rate and includes, but it is not limited to, image, video and language analysis, automated control, and even life science modeling. Besides, deep learning is receiving increasing attention in connection to a vast set of problems in physics where quantitatively accurate outcomes are expected. We consider turbulent velocity signals, spanning decades in Reynolds numbers, which have been generated via shell models for the turbulent energy cascade. Given the multi-scale nature of the turbulent signals, we focus on the fundamental question of whether a deep neural network (DNN) is capable of learning, after supervised training with very high statistics, feature extractors to address and distinguish intermittent and multi-scale signals. Can the DNN measure the Reynolds number of the signals? Which feature is the DNN learning?
Chapter 2: Thermally driven turbulent flows are common in nature and in industrial applications. The presence of a (turbulent) flow can greatly enhance the heat transfer with respect to its conductive value. It is therefore extremely important -in fundamental and applied perspective- to understand if and how it is possible to control the heat transfer in thermally driven flows. In this work, we aim at maintaining a Rayleigh–Bénard convection (RBC) cell in its conductive state beyond the critical Rayleigh number for the onset of convection. We specifically consider controls based on local modifications of the boundary temperature (fluctuations). We take advantage of recent developments in Artificial Intelligence and Reinforcement Learning (RL) to find -automatically- efficient non-linear control strategies. We train RL agents via parallel, GPU-based, 2D lattice Boltzmann simulations. Trained RL agents are capable of increasing the critical Rayleigh number of a factor 3 in comparison with state-of-the-art linear control approaches. Moreover, we observe that control agents are able to significantly reduce the convective flow also when the conductive state is unobtainable. This is achieved by finding and inducing complex flow fields.
|Martin Lellep (Philipps University of Marburg, Germany)||Title: Predictions of relaminarisation in turbulent shear flows using deep learning
Abstract: Given the increasing performance of deep learning algorithms in tasks such as classification during the last years and the vast amount of data that can be generated in turbulence research, I present one application of deep learning to fluid dynamics in this talk. We train a deep learning machine learning model to classify if turbulent shear flow becomes laminar a certain amount of time steps ahead in the future. Prior to this, we use a 2D toy example to develop an understanding how the performance of the deep learning algorithm depends on hyper parameters and how to understand the errors. The performance of both algorithms is high and therefore opens up further steps towards the interpretation of the results in future work.
Pierce Hall 209, 29 Oxford Street
|Detlef Lohse (University of Twente)||Title: Rayleigh vs. Marangoni Abstract: In this talk I will show several examples of an interesting and surprising competition between buoyancy and Marangoni forces. First, I will introduce the audience to the jumping oil droplet – and its sudden death – in a density stratified liquid consisting of water in the bottom and ethanol in the top : After sinking for about a minute, before reaching the equilibrium the droplet suddenly jumps up thanks to the Marangoni forces. This phenomenon repeats about 30-50 times, before the droplet falls dead all the sudden. We explain this phenomenon and explore the phase space where it occurs.
Next, I will focus on the evaporation of multicomponent droplets, for which the richness of phenomena keeps surprising us. I will show and explain several of such phenomena, namely evaporation-triggered segregation thanks to either weak solutal Marangoni flow or thanks to gravitational effects. The dominance of the latter implies that sessile droplets and pending droplets show very different evaporation behavior, even for Bond number << 1. I will also explain the full phase diagram in the Marangoni number vs Rayleigh number phase space, and show where Rayleigh convections rolls prevail, where Marangoni convection rolls prevail, and where they compete.
The research work shown in this talks combines experiments, numerical simulations, and theory. It has been done by and in collaboration with Yanshen Li, Yaxing Li, and Christian Diddens, and many others.
|11/20/2019||Time: 3:00-3:35 pm
Speaker: Haoran Liu
Title: Applications of Phase Field method: drop impact and multiphase turbulence
Abstract: Will a mosquito survive raindrop collisions? How the bubbles under a ship reduce the drag force? In nature and industry, flows with drops and bubbles exist everywhere. To understand these flows, one of the powerful tools is the direct numerical simulation (DNS). Among all the DNS methods, we choose the Phase Field (PF) method and develop some models based on it to simulate the complicated flows, such as flows with moving contact lines, fluid-structure interaction, ternary fluids and turbulence. In this talk, I will firstly introduce the advantages and disadvantages of PF method. Then, I will show its applications: drop impact on an object, compound droplet dynamics, water entry of an object and multiphase turbulence.
Time: 3:35-4:10 pm
Speaker: Steven Chong
Title: Confined Rayleigh-Bénard, rotating Rayleigh-Bénard, double diffusive convection and quasi-static magnetoconvection: A unifying view on their scalar transport enhancement
Abstract: For Rayleigh-Bénard under geometrical confinement, under rotation or the double diffusive convection with the second scalar component stabilizing the convective flow, they seem to be the three different canonical models in turbulent flow. However, previous research coincidentally reported the scalar transport enhancement in these systems. The results are counter-intuitive because the higher efficiency of scalar transport is bought about by the slower flow. In this talk, I will show you a fundamental and unified perspective on such the global transport behavior observed in the seemingly different systems. We further show that the same view can be applied to the quasi-static magnetoconvection, and indeed the regime with heat transport enhancement has been found. The beauty of physics is to understand the seemingly unrelated phenomena by a simplified concept. Here we provide a simplified and generic view, and this concept could be potentially extended to other situations where the turbulent flow is subjected to an additional stabilization.