Events and Seminars

Learning neural operators accurately, efficiently, reliably, and in one shot

Posted on by

dr. lu

Speaker: Dr. Lu Lu – Yale University
Date: Sep 20, 2024; Time: 2:30 PM Location: PWEB 175

Abstract: As an emerging paradigm in scientific machine learning, deep neural operators pioneered by us can learn nonlinear operators of complex dynamic systems via neural networks. In this talk, I will present the deep operator network (DeepONet) to learn various operators that represent deterministic and stochastic differential equations. I will also present several extensions of DeepONet, such as DeepM&Mnet for multiphysics problems, DeepONet with proper orthogonal decomposition or Fourier decoder layers, MIONet for multiple-input operators, and multifidelity DeepONet. I will demonstrate the effectiveness of DeepONet and its extensions to diverse multiphysics and multiscale problems, such as bubble growth dynamics, high-speed boundary layers, electroconvection, hypersonics, geological carbon sequestration, full waveform inversion, and astrophysics. Deep learning models are usually limited to interpolation scenarios, and I will quantify the extrapolation complexity and develop a complete workflow to address the challenge of extrapolation for deep neural operators. Moreover, I will present the first operator learning method that only requires one PDE solution, i.e., one-shot learning, by introducing a new concept of local solution operator based on the principle of locality of PDEs.

Biographical Sketch: Dr. Lu Lu is an Assistant Professor in the Department of Statistics and Data Science at Yale University. Prior to joining Yale, he was an Assistant Professor in the Department of Chemical and Biomolecular Engineering at University of Pennsylvania from 2021 to 2023, and an Applied Mathematics Instructor in the Department of Mathematics at Massachusetts Institute of Technology from 2020 to 2021. He obtained his Ph.D. degree in Applied Mathematics at Brown University in 2020, master’s degrees in Engineering, Applied Mathematics, and Computer Science at Brown University, and bachelor’s degrees in Mechanical Engineering, Economics, and Computer Science at Tsinghua University in 2013. His current research interest lies in scientific machine learning, including theory, algorithms, software, and its applications to engineering, physical, and biological problems. His broad research interests focus on multiscale modeling and high performance computing for physical and biological systems. He has received the 2022 U.S. Department of Energy Early Career Award, and 2020 Joukowsky Family Foundation Outstanding Dissertation Award of Brown University. He is also an action editor of Journal of Machine Learning.

Recent Progress in Black-Box Function Optimization for Industry Problems

Posted on by

Abstract: One of the most common and important problems in the engineering industry is, arguably, to optimize a black-box expensive-to-evaluate function given a strict budget. The function can represent a real-world experiment or a costly simulation code. Specifically, given a set of potential power plant layouts, how do we find the best layout defined against a set of Quantities of Interest? Given a steam turbine, how do we configure its geometry to achieve the best efficiency? How can we optimize the life of a machine by knowing its design variables and how they connect to damage? Given a set of Computational Fluid Dynamics simulations, can we optimize a blade structure for cooling? The problem of course extends to a broad range of other industries and academia as well. As a different example, borrowed from materials discovery, consider a set of binary alloy lattice points: Which atoms should be placed on said points to discover the ground states? Surely, for all these cases, the faster we achieve high-quality optima (ideally global and robust) in terms of resources, the lower the overall cost.

Towards answering these questions, at General Electric Research our team has developed and maintain an industry-strength Efficient Global Optimization scheme called “Intelligent Design and Analysis of Computer Experiments” (IDACE) which builds on a time-tested Gaussian Process meta-model called Bayesian Hybrid Modeling (BHM) originally built from Kennedy O’Hagan’s work.

Having introduced the BHM/IDACE framework, we present in detail a set of successful BHM/IDACE industry case studies and compare to other optimization approaches such as Genetic Algorithms. Finally, we go on to discuss a range of recent modifications and enhancements to these tools all driven from real-world customer needs.

Short Bio: Jesper Kristensen works as a Lead Engineer in the Probabilistics and Optimization team at General Electric’s (GE) Research Center in upstate New York. The team is managed by Dr. Liping Wang. He joined the team in the fall of 2015 as a research engineer. Among other projects, he is currently in charge of a $1MM project leading six engineers to ensure GE stays ahead in Probabilistic capabilities including, but not limited to, meta-modeling, optimization, uncertainty quantification, and uncertainty propagation. He is also the project leader on multiple damage modeling efforts to create Digital Twins of steam turbines.

 

Jesper is a graduate of the Technical University of Denmark (DTU) and holds a Ph.D. from Cornell University in Applied and Engineering Physics advised by Prof. Nicholas Zabaras. His work has generally focused on surrogate modeling and on advanced optimization methods such as adaptive sequential Monte Carlo and Bayesian Global Optimization for improving materials discovery and test cost reduction.

Layer-to-Layer Control in Laser Metal Direct Energy Deposition Additive Manufacturing

Posted on by

Abstract: ​Additive manufacturing (AM), or 3D printing, is beginning to deliver on its long-promised potential to transform industrial production.  Already, tooling and molds are making regular use of AM’s rapid CAD-to-part flexibility to deliver in days what previously took months.  In addition, AM facilitates much greater geometric complexity, which increases the value proposition for AM fabricated parts that are serving in increasingly critical roles.  However, the rate of industrial insertion remains slow due to stubborn problems in process variability arising from the spatial and dynamic complexity of AM, amplifying challenges in qualification.  In-process measurement and analysis, and the utilization of that data in closed-loop feedback control, are widely regarded as the remedy.   This talk will explore one such instance in a blown-powder, direct energy deposition (sometimes referred to as LENS) process.  Here, a laser scanner is used to detect and correct geometric anomalies.  The talk will consider how in-layer and layer-to-layer dynamics may couple to create multi-dimensional dynamic behavior not typically considered, and how novel control methods may stabilize these processes.

Biographical Sketch: Dr. Douglas A. Bristow is currently an Associate Professor in the Department of Mechanical and Aerospace Engineering at the Missouri University of Science and Technology (Missouri S&T).  He received his B.S. in Mechanical Engineering from Missouri S&T in 2001.  He received his M.S. and Ph.D., also in Mechanical Engineering, from the University of Illinois at Urbana-Champaign in 2003 and 2007, respectively.  Dr. Bristow is the Director of the Center for Aerospace Manufacturing Technologies, an industry consortium that currently includes eleven member companies.  He has more than 80 peer-reviewed publications and his research interests include precision motion control, repetitive and iterative process control, additive manufacturing process control, atomic force microscopy, and volumetric error compensation in machine tools and robotics.  Dr. Bristow’s research is currently funded by the National Science Foundation, the Department of Energy, the Digital Manufacturing and Design Innovation Institute, and multiple companies.  He is an Associate Editor at the ASME Journal of Dynamic Systems, Measurement and Control.

Mechanics under the Fold: How Origami Creates Sophisticated Mechanical Properties

Posted on by

Abstract: ​Origami, the ancient Japanese art of paper folding, is not only an inspiring technique to create sophisticated shapes, but also a surprisingly powerful method to induce nonlinear mechanical properties. Over the last decade, advances in crease design, mechanics modeling, and scalable fabrication have fostered the rapid emergence of architected origami structure and material systems. They typically consist of folded origami sheets or modules with intricate three-dimensional geometries, and feature many unique and desirable mechanical properties like auxetics, tunable nonlinear stiffness, multi-stability, and impact absorption. Rich designs in origami offer great freedom to prescribe the performance of such origami structures and materials. In addition, folding offers a unique opportunity of fabrication at vastly different sizes. This talk will highlight our recent studies on the different aspects of origami-based structures and materials–geometric design, mechanics analysis, and achieved properties–and discusses the challenges ahead.

Bio Sketch: Dr. Suyi Li is an assistant professor of mechanical engineering at the Clemson University. He received his Ph.D. at University of Michigan in 2014. After spending two additional years at Michigan as a postdoctoral research fellow, he moved to Clemson in 2016 and established a research group on dynamic matters. His technical interests are in origami-inspired adaptive structures, multi-functional mechanical metamaterials, and bio-inspired robotics. Within his first three years at Clemson, Dr. Li has secured more than one million dollars of research funding, including the prestigious NSF CAREER award. His paper on fluidic origami received the Best Paper Award by the ASME Branch of Adaptive Structures and Material Systems.

A Multiscale Moving Contact Line Theory and Simulation of Droplet Spreading and Cell Durotaxi

Posted on by

Abstract: In this talk, we present a novel multiscale moving contact line (MMCL) theory, which offers a powerful numerical simulation method for modeling and analysis of dynamic wetting, liquid droplet spreading on solid substrates, and various capillary motion phenomena. In the proposed multiscale moving contact line theory, we couple molecular scale adhesive interaction i.e. the van der Waals type interaction force and the macroscale fluid mechanics to solve droplet motions on solid substrates. In specific, we combine a coarse-grained adhesive contact model with a modified Gurtin-Murdoch surface hydroelasto-dynamics theory and the Navier-Stokes equation in the bulk fluids to formulate the multiscale moving contact line hydrodynamics theory in order to simulate a broader class of colloidal and soft matter physics phenomena, and related chemomechanical problems, such as cell motility, water spider walking, colloid suspension, and gas bubble in water, etc.

The advantage of adopting the coarse grain adhesive contact model in the moving contact line theory is that it can levitate and separate the liquid droplet with the solid substrate, so that the proposed multiscale moving contact line theory avoids imposing the non-slip condition, and then it removes the subsequent shear stress singularity problem, which allows the surface energy difference and surface stress propelling droplet spreading naturally.

We have also developed a soft matter model for biological cells that can model actin polymerization and ATP hydrolysis, and retrograde flow in cellular lamellipodia. By employing the MMCL method, we have successfully simulated cell durotaxi over the soft elastic substrates with non-uniform elastic stiffness.  By employing the proposed method, we have successfully simulated droplet spreading over various elastic substrates and cell durotaxi over the substrates with non-uniform elastic stiffness. The obtained numerical simulation results compare well with the experimental and molecular dynamics results reported in the literature.

Biographical Sketch:  Dr. Shaofan Li is currently a full professor of applied and computational mechanics at the University of California-Berkeley. Dr. Li graduated from the East China University of Science and Technology (Shanghai, China) with a BS degree in 1982; he also holds MS Degrees from both the Huazhong University of Science and Technology (Wuhan, China) and the University of Florida (Gainesville, FL, USA) in 1989 and 1993 respectively. In 1997, Dr. Li received a PhD degree from the Northwestern University (Evanston, IL, USA), and he was also a post-doctoral researcher at the Northwestern University during 1997-2000. In 2000, Dr. Li joined the faculty of the Department of Civil and Environmental Engineering at the University of California-Berkeley. Dr. Shaofan Li is the recipient of IACM Fellow Award [2017]; Distinguished Fellow Award of ICCES [2014]; ICACM Computational Mechanics Award [2013], USACM Fellow Award (2013), A. Richard Newton Research Breakthrough Award [2008], and NSF Career Award [2003]. Dr. Li has published more than140 articles in peer-reviewed scientific journals (SCI) with h-index 43 (Google Scholar), and he is also the author of two research monographs/graduate textbooks.

Predicting Fuel Properties of Potential Biofuels Using an Improved Artificial Neural Network Based on Molecular Structure

Posted on by

Abstract: The next generation of alternative fuels is being investigated through advanced chemical and biological production techniques for the purpose of finding suitable replacements to diesel and gasoline while lowering production costs and increasing process yields. Chemical conversion of biomass to fuels provides a plethora of pathways with a variety of fuel molecules, both novel and traditional, which may be targeted. In the search for new fuels, an initial, intuition-driven prediction of fuel compounds with desired properties is required. Due to the high cost and significant production time needed to synthesize these materials for testing, a predictive model would allow chemists to screen fuel properties of potentially desirable fuel candidates at the ideation stage. Recent work has shown that predictive models, in this case artificial neural networks (ANN’s) analyzing quantitative structure property relationships (QSPR’s), can predict the cetane number (CN) of a proposed fuel molecule with relatively small error. A fuel’s CN is a measure of its ignition quality, typically defined using prescribed ASTM standards and a cetane testing engine. Alternatively, the analogous derived cetane number (DCN), obtained using an Ignition Quality Tester (IQT), is a direct measurement alternative to the CN that uses an empirical inverse relationship to the ignition delay found in the constant volume combustion chamber apparatus. Model validation and expansion of the experimental database used in this study implemented DCN data acquired using an IQT. The present work improves on an existing model by optimizing the model architecture along with the key learning variables of the ANN and by making the model more generalizable to a wider variety of fuel candidate types. The approach enables researchers to focus on promising molecules by eliminating less favorable candidates in relation to their ignition quality.

Biographical Sketch: Hunter Mack is an Assistant Professor in the Department of Mechanical Engineering at the University of Massachusetts Lowell.  His research focuses on combustion, biofuels, and energy efficiency.  Prior to joining UML, he was a Project Scientist & Lecturer at the University of California at Berkeley, a Senior Engineer at solar concentrator start-up Banyan Energy, and a Postdoctoral Researcher in the Combustion Analysis Laboratory at UC Berkeley.  He received his M.S. (2005) and Ph.D. (2007) from UC Berkeley with an emphasis on multi-component fuels in Homogeneous Charge Compression Ignition (HCCI) engines. He also holds a B.S. in Mechanical Engineering from Washington University in St. Louis and a B.A. in Physics from Hendrix College (Conway, Arkansas).

Active Thermosyphons to Condense Water from Power Plant Flue Gas, and an Overview of Energy Research at Stony Brook University

Posted on by

The talk will focus on two themes.  The first will be an overview of our current ARPA-E project to condense water from flue gas for dry-cooled power plants power plants. Water use by power plants is an increasing concern across the U.S., and is particularly problematic in arid regions, such as the southwest.  In this project, an advanced two-phase thermosyphon concept is employed that removes several of the traditional limitations of conventional thermosyphons. The condensed water can be used to pre-cool the condenser air to reduce the effective ambient temperature, for turbine inlet air evaporative cooling, or for other uses in the plant, as needed. The second portion of the talk will provide a short overview of other energy-related research activities in the Mechanical Engineering department at Stony Brook University with the goal to explore future opportunities for collaboration and joint projects between both institutions.

Bio: Jon Longtin joined the Mechanical Engineering Faculty at Stony Brook University in 1996. He came to Stony Brook after receiving his Ph.D. degree in 1995 from U.C. Berkeley, followed by a one-year postdoc at the Tokyo Institute of Technology in Japan. His research interests include energy conservation, innovative energy transfer and storage, and energy monitoring and diagnostics, as well as laser materials processing, particularly with ultrafast lasers and the development of sensors for harsh environments. His research has been funded by NSF, DOE, DOD, NASA, NYSERDA, and a variety of industrial sources.  He is the author of over 140 technical publications and holds 11 issued and pending patents. He has received the Presidential Early Career Award for Scientists and Engineers, two Excellence in Teaching Awards, and an R&D 100 award.  He is a licensed Professional Engineer in New York State and serves as a technical advisor to a variety of companies and non-profit organizations.

Molecular simulations of mechanical properties for polymer materials

Posted on by

The U.S. Army Research Laboratory (ARL) was activated 25 years ago with a mission to discover, innovate and transition science and technology to ensure dominant strategic land power. One of key research strategies at ARL is a development of superior protection systems for individual warfigter and vehicles. The protective systems often use polymers due to their low weight, good strength and toughness which improves resistance to ballistic penetration.

In this presentation, we will review fundamental research projects carried out at ARL and aimed at understanding and prediction of mechanical properties for polymers at high strain rates. We will focus on semicrystalline polyethylene discussing  onset of crystallization in polymer melt upon drawing and cooling. Our model includes amorphous domains, explains experimentally known mechanics and supports fracture mechanism through chain pulling. We will discuss shock propagation in anisotropic semicrystalline models at atomistic level.

We also study emerging 2D polymers inspired by Kevlar(R) which is well known for its remarkable strength and stiffness facilitated by the hydrogen bonds formed between Kevlar chains. Molecular and micromechanical calculations predict that ensembles of 2D molecules bonded with hydrogen bonds will form stiff, strong and tough films of unprecedented mechanical performance.

Bio: Dr.  Jan Andzelm serves as the Team Leader of the Multiscale Modeling Team in the Polymer Branch at ARL investigating properties of materials important for Soldier protection. The team is developing and applying novel computational techniques at quantum mechanical, atomistic and mesoscale levels aimed at understanding and predicting structural, mechanical and electronic properties of macromolecules and composite materials.

Dr. Andzelm has co-authored over 140 scientific papers and book contributions that attracted more than 14000 citations and generated h-index of 42. Dr. Andzelm was selected as a U.S. Army Research Laboratory Fellow in 2010.

Isogeometric Methods for Solids, Structures, and Fluid-Structure Interaction: From Early Results to Recent Developments

Posted on by

Abstract: This presentation is focused on Isogeometric Analysis (IGA) with applications to solids and structures, starting with early developments and results, and transitioning to more recent work. Novel IGA-based thin-shell formulations are discussed, and applications to progressive damage modeling in composite laminates due to low-velocity impact and their residual-strength prediction are shown. Fluid–structure interaction (FSI) employing IGA is also discussed, and a novel framework for air-blast-structure interaction (ABSI) based on an immersed approach coupling IGA and RKPM-based Meshfree methods is presented and verified on a set of challenging examples. The presentation is infused with examples that highlight effective uses of IGA in advanced engineering applications.

Bio: Yuri Bazilevs is the E. Paul Sorensen Chair in the School of Engineering at Brown University. He was previously a Professor and Vice Chair in the Structural Engineering Department at the University of California, San Diego. Yuri is the original developer of Isogeometric Analysis (IGA), a new computational methodology that aims to integrate engineering design (CAD) and simulation (FEM). For his research contributions Yuri received a number of awards and honors, including the 2018 ASCE Walter L. Huber Research Prize. He is included in the 2014-2018 lists of Highly Cited Researchers, both in the Engineering and Computer Science categories.