To read the speakers abstract please click on the link below associated with their name.


Plenary Speakers

Title: Artificial Intelligence (AI) for High Performance Computing (HPC) Analytics

Speaker: Ayse Kivilcim Coskun, Department of Electrical and Computer Engineering, Boston University

Abstract: Today’s HPC systems face challenges in delivering predictable performance, while maintaining efficiency, resilience, and security. Much of computer system management has traditionally relied on (manual) expert analysis and policies that rely on heuristics derived based on such analysis. This talk will discuss a new path on designing “automated analytics” methods for HPC systems and how to make strides towards a longer term vision where computing systems are able to self-manage and improve. Specifically, the talk will first cover how to systematically diagnose root causes of performance “anomalies”, which cause substantial efficiency losses and higher cost. Second, it will discuss how to identify applications running on computing systems and discuss how such discoveries can help reduce vulnerabilities and avoid unwanted applications. The talk will also highlight how to apply machine learning in a practical and scalable way to help understand complex systems, demonstrate methods to help standardize study of performance anomalies, and point out future directions in automating HPC system management.


Title: Sustainable AI: Environmental Implications, Challenges and Opportunities

Speaker: Carole-Jean Wu, Research Scientist and Technical Lead Manager at Meta

Abstract: The past decade has witnessed orders-of-magnitude increase in the amount of compute for AI and emerging use cases. Modern natural language processing models are fueled with over trillion parameters while the memory needs of deep learning recommendation and ranking models have grown from hundreds of gigabytes to the terabyte scale. We will explore the environmental implications of the super-linear growth trend for AI from a holistic perspective, spanning data, algorithms, and system hardware. I will talk about the carbon footprint of AI computing by examining the model development cycle across industry-scale use cases and, at the same time, considering the life cycle of system hardware. The talk will capture the operational and manufacturing carbon footprint of AI and computing. I will present an end-to-end analysis for what and how hardware-software design and at-scale optimization can help reduce the overall carbon footprint of AI and computing. Based on the industry experience and lessons learned, I will share the key challenges across the many dimensions of AI. This talk will conclude with important development and research directions to advance the field of AI and for digital technologies in an environmentally-responsible and sustainable manner.

Invited Speakers

Title: Hidden Masses: The next frontier in observational gravitational-wave astronomy

Speaker: Sarah Caudill, Department of Physics, UMass Dartmouth

Abstract: The last seven years have seen the explosive start of the field of observational gravitational-wave (GW) science. Since the first detection of a GW signal from a binary black hole merger in 2015 by the advanced detectors of the Laser Interferometer Gravitational-wave Observatory (LIGO), more than 100 new signals have been detected by LIGO and the European Virgo detector. Added to these is the spectacular low-latency discovery of GWs from a binary neutron star system, which ushered in the era of multi-messenger astronomy with GWs. The science enabled by these detections has been immense, from precision tests of General Relativity to independent measurements of the Hubble constant. However, we have not yet begun to access the full, rich discovery space of GW signals. In this talk, I will discuss our plans to reveal the next frontier in GW detections, from compact binary systems with complex dynamics to exotic systems with the lightest (and heaviest) black holes that can be detected by LIGO and Virgo. I will also discuss the computing challenges our field is currently facing and our move toward domain-specific computing for the third-generation of gravitational-wave detectors.

Title: Investigations of Electric Field Effects on Catalysis: A Combination of Deep Learning Models and Multi-Scale Simulations

Speaker: Fanglin Che, Department of Chemical Engineering, UMass Lowell

Abstract: External electric fields can modify binding energies of reactive surface species and enhance catalytic performance of heterogeneously catalyzed reactions. Large electric fields can be experimentally generated through three ways:

  1. internally over molecular length scales in (metallo-)enzyme and zeolite catalytic active sites;
  2. externally in gas / solid heterogeneous catalytic system, such as ultra-high vacuum conditions via scanning tunneling microscopy, field ion / emission microscopy, or flow reactor type via dielectric barrier discharge, coaxial capacitor, and microwave reactor; and
  3. in an interfacial way at gas / liquid / solid triple phase boundary.

Compared to experimental studies, theoretical work on electric field effects in catalysis is very limited due to the low efficiency of pure Density Functional Theory (DFT) calculations for predicting field-dependent energetics of catalytic reactions. This has led to an incomplete picture of how electric fields influence catalytic mechanisms at the atomic-scale and hinders the design and optimization of field-induced catalytic technologies.

To address this gap, we designed a novel deep learning framework for predicting the field-dependent adsorption energies. Specially, we employed a Graph Neural Network (GNN) to capture the relationship among the geometries, followed by a shared multiple-layer perception (MLP) for field-induced catalytic reaction prediction. The deep learning algorithm developed here accelerates field-dependent energy predictions with acceptable accuracies by five orders of magnitudes compared to DFT alone and has the capacity of transferability, which can predict field-dependent energetics of other catalytic surfaces with high quality performance using little training data. Our designed deep learning framework can provide potential good catalyst candidates for field-induced heterogeneous catalysis in a short time. By this means, some unacceptable catalyst candidates can be quickly filtered out, thus avoiding the unnecessary computations.

Title: Coastal Hazards, Analysis, Modeling, and Prediction for Emergency Management and Response

Speaker: Isaac Ginis, Graduate School of Oceanography, University of Rhode Island (URI) and a team of URI researchers

Abstract: Emergency managers need access to relevant, local-scale, actionable information about the potential consequences of extreme events in advance of a storm's landfall. The Rhode Island Coastal Hazards, Analysis, Modeling, and Prediction (RI-CHAMP) system advances storm modeling and simulation capabilities through a near-real-time hazard and impact prediction system for hurricanes and nor'easters in Southern New England. RI-CHAMP integrates end-user concerns regarding storm impacts with the ADvanced CIRCulation (ADCIRC) model wind, wave, and storm surge outputs to provide predictions of the consequences of extreme weather impacts on critical infrastructure (e.g., wastewater treatment facilities, hospitals, police stations, and seaports). RI-CHAMP can be used for long-term planning for critical infrastructure resilience and emergency response operations using near-real-time storm model predictions. This project engages key end-users in developing and disseminating the tools to make them relevant and useful in planning and response, including federal, state, and local partners.

Title: High-fidelity multiphase flow simulations using HPC

Speaker: Aswin Gnanaskandan, Department of Mechanical Engineering, Worcester Polytechnic Institute (WPI)

Abstract: Multiphase flows play an important role in aerospace, marine and biomedical applications with both beneficial (e.g. atomization of fuel, lithotripsy) and detrimental effects (e.g. erosion, loss of efficiency). The ability to accurately predict multiphase flow behavior is imperative to control its formation, exploit its benefits and reduce its detrimental effects. A major difficulty in accurately predicting multiphase flows lies in the complex multiple length scale nature of such flows. Applying a universal model to cater to all the length scales is prohibitive in terms of computational cost. Hence, the most practical way to solve the multiscale problem is to develop models separately at both micro and macro length scales and then develop appropriate bridging models that can transition smoothly between models at different length scales. In this talk, I will discuss my research on two important areas of multiphase flow modeling in the realms of macroscopic and microscopic length scales which are the building blocks of a multiscale model and the importance of HPC in these evaluating these models.

First, I will discuss my work in the area of cavitation, a physical phenomenon that leads to gas phase formation in a liquid due to a reduction in pressure. I will present the development and application of a Large Eddy Simulation (LES) methodology for cavitating flows and discuss the importance of non-dissipative numerical methods and turbulence modeling in the prediction of cavitation. The enhanced accuracy of LES compared to low fidelity methods in modeling complex cavitating flows will be demonstrated. Under the next topic, I will discuss my research on the development of a Euler-Lagrange based simulation methodology for modeling gaseous micro bubbles dispersed in a liquid/solid medium. An important biomedical application of this method is the modeling of High Intensity Focused Ultrasound (HIFU), a non-invasive therapy for cancer treatment.

Title: Processing of Massive Biological Datasets at UMass Boston

Speaker: Daniel Haehn, Department of Computer Science, UMass Boston

Abstract: We are using the Chimera cluster at UMass Boston to (among others) create the world's largest publicly available breast cancer database and visualize huge electron microscope images of the fruitfly retina. For this, we used Docker in a Slurm environment, found practical and not-so-practical ways for external data transfers, discovered the best coding setups to access GPUs, and in general, pushed our sysadmins and RC support staff to the absolute limits.

Title: Molecular Simulation of Deep Eutectic Solvents and Lipid Nanoparticles for Gas Separations and Drug Delivery

Speaker: Francisco Hung, Department of Chemical Engineering, Northeastern University

Abstract: In this talk, I will present our recent efforts on using classical molecular dynamics simulations on two independent projects. I will first give an overview of our work on modeling deep eutectic solvents (DESs) inside nanopores, for applications in separations of CO2 from methane. DESs result from mixing a quaternary ammonium salt such as choline chloride with a hydrogen bond donor species such as urea, which forms a eutectic mixture with much lower melting point compared to the pure components. DESs share many of the remarkable properties of ionic liquids but are significantly cheaper, biodegradable and more environmentally friendly. Molecular dynamics simulations are used to fundamentally understand the properties of different DESs inside nanopores of different sizes and chemical nature, for applications in separations of CO2 from methane. In the second part of this talk, I will provide an overview of our work on modeling novel formulations of hyper-elastic liposomes, which have been shown in preliminary experimental studies to have larger in vivo accumulation in tumors, and thus could be more effective in delivering anti-cancer drugs compared to stiffer nanoparticles. Using molecular simulations, we aim at fundamentally understanding how the mechanical properties of the liposomes are affected by the molecular structure of the lipids and the composition of the liposome.

Title: Information divergences and optimal transport for enhanced generative modeling

Speaker: Markos Katsoulakis, Department of Mathematics and Statistics, UMass Amherst

Abstract: We present recent work on variational representations for probability divergences and metrics with applications to machine learning and uncertainty quantification (UQ). The newly constructed information-theoretic divergences interpolate between f-divergences (e.g. KL-divergence) and Integral Probability Metrics (IPM) such as the Wasserstein or the MMD distances.

These divergences show improved convergence and stability properties in statistical learning applications (in particular for generative adversarial networks (GANs)) as well as tighter uncertainty regions in UQ.

These divergences also provide new mathematical and computational insights on Lipschitz regularization methods (e.g. spectral normalization in neural networks) which have recently emerged as an important algorithmic tool in Deep Learning. A version of the Data Processing Inequality allows flexibility in selecting the functions to be optimized over in the variational representation of the divergences.

This feature comes in particularly handy when learning distributions which preserve additional structure such as group symmetries, or more generally geometric constraints.

Combining our new divergences with recent advances in invariant and equivariant neural networks allowed us to introduce Structure-Preserving GANs (SP-GAN) as a data-efficient approach for learning distributions with symmetries.

Our theoretical insights lead to a reduced invariant discriminator space, as well as to carefully constructed equivariant generators, avoiding flawed designs that can easily lead to a catastrophic “mode collapse” for the learned distribution.

Our experimental results show a drastic improvement in sample fidelity and diversity, and importantly in the amount of data needed to learn invariant distributions.

Title: From nucleotides to neural networks: studying gene regulation in the era of “big data”

Speaker: Jill Moore, Bioinformatics and Integrative Biology, UMass Chan Medical School

Abstract: The last decade has witnessed tremendous success in discovering molecular foundations of human diseases using high throughput sequencing experiments and large-scale data analysis. Studies have profiled the genomes, epigenomes, and transcriptomes of hundreds of cell and tissue types generating terabytes of sequenced nucleotides. While individual datasets can be used to study human biology, when integrated, the wealth of data collectively provides a far richer, much more comprehensive picture of gene regulation and genome function. However, such integrations require references to anchor complex analyses, which do not exist for many components of gene regulation. Here we introduce the Registry of candidate cis-regulatory elements (cCREs), a collection of over one million putative regulatory regions in the human genome curated by integrating thousands of multi-omic sequencing experiments. These cCREs span hundreds of unique cell and tissue types and shed insight onto fundamental gene regulation principles. Additionally, the Registry of cCREs enables us to better understand the relationship between DNA sequence and element function through deep learning frameworks, demonstrating that specific sequence combinations are associated with transcriptional activity. Overall the Registry of cCREs provides an invaluable framework for studying gene regulation and better understand diseases.