Browsing by Subject "high performance computing"

Now showing 1 - 20 of 40

Accelerating Discoveries with Large Language Models
(2023-09-29) Kumar, Krishna
Coupled Magnetic Domain Wall Oscillations for Neuromorphic Lateral Excitation Behavior in Periodic Ferromagnetic Nanowire Arrays
(2023-09-28) Park, Jiwoo; Rogers, Vivian; Incorvia, Jean Anne
"The Domain Wall (DW) racetrack is a ferromagnetic nanowire that encodes data in the spatial position of magnetization boundary—a magnetic DW. Many emerging spintronic nanodevices utilize DW racetracks for post-Von-Neumann computing applications, as nonvolatile analog information encoded in DW position can be modified and later decoded by magnetic tunnel junctions (MTJ). We explore the physics of DW oscillations for use in neuromorphic computing, an emerging computer architecture inspired by biological spiking neural systems. In neurons, lateral excitation and resonant firing are behaviors where neurons excite their neighbors without direct connections or spike when driven at a certain frequency. Observing these behaviors in DW racetracks is of great value to us in terms of power-efficiency and computing capability, by natively encoding neuro-mimetic behaviors in the hardware. To study this, we construct a mathematical model where we treat DW racetracks as damped harmonic oscillators with magnetostatic interactions. Then, a Green’s function analysis is applied to the Hamiltonian of the system, obtaining the transfer function of racetracks responding to an external magnetic field or current. By utilizing the transfer function of a racetrack with the convolution theorem, we estimate any DW position displacement resulting from signal applied to a neighboring racetrack. To test the model, we simulated CoFeB racetracks in the Mumax3 micromagnetic software on TACC’s Lonestar6. As a result, coupled oscillations were observed, implying the possibility of lateral excitation. We predict magnetically co-aligned racetracks would allow neuro-realistic behavior, where a bias current tilt the DW pinning potential so that oscillations allow the DW to de-pin and fire easier; magnetically anti-aligned racetracks would exhibit DW oscillations due to the intrinsic resonance of the magnetic stray field DW interactions. The results show the frequency-dependent behaviors of magnetically coupled DW-MTJ racetracks and their possible application in nanoelectronic filters, antennas, and neuromorphic computing."
Diffusion and NMR Relaxation Properties of Confined Fluids in Organic Nanopores: A Molecular Dynamics Simulation Study
(2023-09-28) Amaro-Estrada, Jorge Ivan; Wang, You; Torres-Verdín, Carlos
In this study, we conducted a series of molecular dynamics (MD) simulations to investigate the impact of pore size, pore shape, and paramagnetic impurities on the diffusion and nuclear magnetic resonance (NMR) properties of confined water and oil within kerogen nanopores. A comprehensive analysis was performed on the mean-squared displacement (MSD), diffusion coefficients (DA), and NMR relaxation times of each fluid under confinement. The results revealed that the presence of kerogen affects the diffusion behavior of water and oil in organic nanopores, with smaller pore sizes leading to reduced mobility (by 1 order of magnitude). At a Larmor frequency of 400 MHz, both n-pentane and water in kerogen pores exhibit longer relaxation times than at lower frequencies (2.3 and 22 MHz). Additionally, the inclusion of paramagnetic impurities and a more realistic pore network in the simulations led to diffusion coefficients and T1 relaxation times comparable to experimental values reported in the literature. Results reported here could contribute to a better understanding of the molecular interactions between the fluids and the nanoporous environment. All MD simulations were performed using the GROMACS software version 2019.6 installed in the Texas Advanced Computing Center (TACC) at The University of Texas at Austin.
Distinguishing Gravitational Wave Parameter Markers: Eccentricity vs Precession
(2023-09-28) Tibrewal, Snehal; Iglesias, Hector; Lange, Jacob; Ferguson, Deborah; Shoemaker, Deirdre
When two massive accelerating objects collide, they generate ripples in the space-time continuum, referred to as gravitational waves (GWs). These GW signals can be detected by the LIGO detectors and provide us with information on the merging binary black holes (BBH). Our work aims to study the distinguishability between eccentricity and spin precession - two important parameters of a GW signal. There has been evidence of ambiguity when attempting a recovery of these parameters due to similarities in their GW markers. Since real data is limited and our knowledge is mostly biased by the accuracy of search pipelines, we are using injections that are signals from simulated BBH mergers. These simulated signals are obtained using numerical relativity (NR) and include both eccentricity and spin precession, something which previous studies haven't been able to mimic due to limitations of current day waveform models. TACC plays a central role in our ability to run these NR simulations. Without supercomputing resources like TACC, the computational complexity of solving Einstein's Field Equations can take months, or even years. We utilize our in-house NR code Maya-Waves to run simulations for various configurations. With these simulations, we are able to control the explorable parameter space as well as test the accuracy of our recovery pipelines. Additionally, these NR simulations will be highly advantageous to the future development of more accurate waveform models. Our preliminary results point towards possible issues with distinguishability, especially in certain parameter spaces. However we continue to run experiments to be able to comment more clearly on these bounds.
Efficient Glitch Detection in Gravitational Wave Data using Particle Swarm Optimization
(2023-09-28) Girgaonkar, Raghav; Mohanty, Soumya
The direct detection of Gravitational Waves (GWs) starting in 2015 has opened a new window on the observable Universe. It has uncovered previously unseen but extremely violent phenomena such as the mergers of Black Holes and Neutron Stars, also known as Compact Binary Coalescences (CBCs). However, the distance to which such CBCs can be detected is currently limited by the large false alarm rate arising from transient interference of instrumental origin, or “glitches”, in gravitational wave data. To address this problem, we propose a novel approach enabled by Particle Swarm Optimization (PSO), a nature-inspired global optimization metaheuristic. PSO offers computational efficiency surpassing traditional grid-based matched-filtering searches, allowing the search for GW signals to be expanded to unphysical sectors of the signal parameter space. While a CBC is physically characterized by the masses of the component stars (among other parameters) the preferred parameters for defining the GW signal from a CBC are the so-called chirp time parameters obtained under a non-linear transformation of the component masses. The chirp time parameter space is larger than the mass space and separates into a physical and unphysical region. We find that searching for the signals across both the regions is a very effective strategy in separating the glitch population from the astrophysical one. We observe that the glitches tend to have unphysical estimated chirp times while the astrophysical signals do not. This novel strategy can address both strong and weak glitches unlike existing strategies that are mostly effective for the former. Our method is demonstrated on 10 hours of real GW data from the LIGO detectors using a pipeline implemented on Lonestar6. On this data, we were able to identify and veto over 98% of the glitches present in GW detector data, offering a robust and efficient solution to enhance GW search sensitivity.
Exploring the Structure of [Ni(dippe)] Fragment via Molecular Mechanical and Quantum Mechanical Calculations
(2023-09-28) Sanchez, Jennifer; Gallegos, Dominique; Atesin, Abdurrahman Cagri; Jones, William D.; Ateşin, Tülay
The [Ni(dippe)] fragment has been successfully used in strong bond activation, such as C—H, C—C and C—S bonds. While the activity of transition metals depends on the nature of the metal, their reactivity is largely determined by the ligands used. Therefore, understanding the metal−ligand structure is crucial for enhancing reactivity and selectivity. To this end, a conformational analysis of the [Ni(dippe)] fragment was conducted using molecular mechanics and quantum mechanics calculations after a thorough benchmark study. A comprehensive systematical analysis of conformers will be presented. The insights gained from this research can be applied to the increasingly important field of Ni-catalyzed reactions.
Flow and Scalar Transfer Characteristics for a Circular Colony of Vegetation
(2022-09-29) Kingora, Kamau; Raza, Mishal; Sadat, Hamid
Local and global flow structures, as well as transfer and transport of a passive scalar from a circular colony of uniformly distributed vegetation stems, are investigated at Re = 2100, Re = 4200, and Re = 8400. The number of stems in the colony is varied from 1 to 284 yielding a solid fraction of 0.0<𝜙𝜙<0.65. The following three flow regimes are identified: a co-shedding flow regime prevails at low solid fraction where wakes of individual cylinders have minimal interaction; a bleeding-wake flow regime is identified at intermediate solid fraction in which stream-wise bleeding flow delays the formation of colony-scale vortices yielding a steady wake between two separated shear layers; and a single-body flow regime is observed at high solid fraction and is accompanied by the commencement of colony-scale vortex shedding. As Reynolds number increases, the separated shear layers observed at intermediate solid fraction break up to form stem- scale vortices that organize themselves in colony scale coherent structures. As the solid fraction increases, drag and Sherwood number experienced by colonies increases linearly and at a reducing rate at low and intermediate solid fractions, respectively, while the net lift remains negligible. At high solid fraction, the commencement of colony-scale vortex shedding is accompanied by a jump in lift and base suction. Pressure and friction lift/drag increase and decrease with an increase in solid fraction, respectively, toward the value experienced by a solid cylinder. Sherwood number, on the other hand, decays exponentially toward the value experienced by a solid cylinder at high solid fraction. Colonies at intermediate solid fraction exhibit the highest scalar transfer but weakest transport in their near field wake. Scalar transfer in colonies with high solid fraction deteriorates with an increase in solid fraction yielding less scalar concentration in their downstream wake. Each case consist of about 14M computational points and computations were performed on TACC LS6 clusters. A typical case converges in 128,000 processor hours.
GPU-Acceleration of PDE Solvers for Large-Scale Wave Simulation
(2022-09-29) Hanindhito, Bagus; Gourounas, Dimitrios; Fathi, Arash; Trenev, Dimitar; Gerstlauer, Andreas; John, Lizy K.
Large-scale simulations of wave-type equations have many industrial applications, such as in oil and gas exploration. Realistic simulations, which involve a vast amount of data, are often performed on multiple nodes of an HPC cluster. Using GPUs for these simulations is attractive due to considerable parallelizability of the algorithms. Many industry-relevant simulations have characteristics in their physics or geometry that can be exploited to improve computational efficiency. Furthermore, the choice of simulation algorithm impacts computational efficiency significantly. In this work, we exploit these features to significantly improve performance for a class of problems. Specifically, we use the discontinuous Galerkin (DG) finite element method, along with the Gauss- Lobatto-Legendre (GLL) integration scheme on hexahedral elements with straight faces, which then greatly reduces the number of BLAS operations, and simplify the computations to Level-1 BLAS operations, reducing the turnaround time for wave simulation. However, attaining peak performance of GPUs is often not possible in these codes that exacerbate bottlenecks caused by data movement, even when modern GPUs enjoying the latest high-bandwidth memory are being used. We have developed an efficient and scalable GPU-accelerated PDE solver for Wave Simulation, by using hardware- and data-movement-aware algorithms. While significant speed-up over CPUs can be achieved, data movement still limits GPU performance. We present several optimization strategies, including kernel fusion, Look-Up-Table-based neighbor search, improved shared memory utilization, and SM-occupancy-aware register allocation. They improve performance up to 84.15x over CPU implementations and 1.84x over base GPU implementations on average. We then extend the functionality to support multi-GPUs on multi-node HPC clusters for large-scale wave simulations, and perform additional optimizations to reduce communication overhead. We also investigate the performance of several MPI libraries in order to fully overlap communication and computation. We are able to reduce the communication overhead by 70%, and achieve weak-scaling over 128 GPUs in TACC Longhorn GPU cluster.
Graph Theory Modeling on the BrainMap Community Portal: Network topology reveals a central role for the medial frontal gyrus in mesial temporal lobe epilepsy
(2023-09-28) Towne, Jonathan M.; Eslami, Vahid; Cavazos, José E.; Fox, Peter T.
"Mesial temporal lobe epilepsy(MTLE) is a disorder of neural networks, often amenable to surgical treatment. Yet, resection of the seizure-onset zone can be non-curative. Seizure recurrence is attributable to distributed network-mediation of ictogenesis(seizure-onset). To address reasons for surgical failure, it is crucial to understand higher-level network properties in MTLE pathology. Graph theory models(GTM) interrogate network properties by quantifying features of network topology. Imaging studies have leveraged graph theory to detect compensatory network changes and predict seizure-onset laterality in MTLE, portending GTM utility in biomarker development. GTM was recently adapted for coordinate-based meta-analysis(CBMA) of voxel-based morphometry(VBM) and physiology(VBP) studies, instantiating a multi-variate extension of activation likelihood estimation(mass-univariate-CBMA). We applied meta-analytic-GTM(M-GTM) to VBM/VBP-studies of MTLE, to infer organizational properties of MTLE network pathology. To quantify topological organization of MTLE network pathology, BrainMap applications (M-GTM/Mango) were used to derive a co-alteration GTM from 74 experiments. M-GTM was used to model coordinates of MTLE pathology as spatial probability distributions, defining nodes at peaks in the joint distribution of pathology and computing edges as co-alterations in individual studies. Clusters of network pathology (modules) were detected via spectral-partition and interpreted in Mango via Regional Behavioral/Diseases Analyses of the BrainMap database(21,435-Task-Activation/4,398-VBM-experiments;n=107,167/115,627-subjects). MATLAB/Cytoscape were used to compute topology metrics and nodal influence on MTLE network topology. Two distributed network-modules were identified in the MTLE co-alteration network, connected only via three most-influential nodes: hippocampus/MDN-thalamus/medial frontal gyrus. Module-1 regions(M1:mesial-temporal/deep-nuclear/frontal/precentral/postcentral/cingulate/inf-parietal structures) were associated with emotion-cognition(Z=3.7) and weakly with social-cognition/explicit-memory(Z=2.5/2.3). Module-2(M2:cerebellar/occipital/temporoparietal/precentral/inf.frontal/med.frontal gyri) regional associations included language/speech/semantic-cognition(Z=4.1/3.0/2.3). M1 matched known VBM-patterns in Alzheimer’s(Z=4.1); both M1/M2 matched those of structural epilepsy pathology(Z=3.4/3.6). A 2-node module(M3:parahippocampus/amygdala) and disconnected pair(M4:mid.frontal/cingulate gyri) were noted. Discrete co-alteration networks exist in MTLE. The medial frontal gyrus likely mediates interactions and evolution of limbic-M1 and verbal-M2 symptoms in MTLE. Pathology modules and intermodular connections represent potential targets for disease monitoring/therapeutic modulation. This study was funded by R01MH074457, T32GM113896, T32TR004545, F31NS131025, and the American Epilepsy Society."
How To Incorporate CI/CD When Your Science Requires HPC
(2022-09-29) Stuart, Georgia
Impact of Tricuspid Annuloplasty Device Shape and Size on Valve Mechanics: A Virtual Case Study
(2023-09-28) Haese, Collin E.; Mathur, Mrudang; Rausch, Manuel K.
"Tricuspid valve disease affects 1.6 million Americans [1]. The primary surgical treatment for tricuspid valve disease is the implantation of annuloplasty devices – ring like devices which come in various shapes and sizes. However, selection of ring size and shape are often motivated by surgeon preference rather than scientific rationale. We used our subject-specific finite element model of the human tricuspid valve, the Texas TriValve 1.0 [2], to conduct a virtual case study in order to understand the impact of device size and shape on valve mechanics and provide a rational basis for device selection. To this end, we implanted four different annuloplasty devices of six different sizes in our virtual patient. All finite element simulations were solved using the Texas Advance Computing Center’s Stampede2 supercomputer. After each virtual surgery, we computed the coaptation area, leaflet end-systolic angles, leaflet stress, and chordal forces. In Figure 1 we see the results of a baseline simulation of our healthy and diseased Texas TriValve. Figure 2 shows the outcome of all 24 virtual repair cases. Our results showed the choice of device shape and size significantly impacts valve mechanics. We found that the one flat device, the Edwards Classic, maximized coaptation area and minimized leaflet stress and chordal forces while the contoured devices were better at normalizing end-systolic angles. Further, we found reducing device size increased coaptation area while negatively impacting stress, chordal forces, and end-systolic angles. Our case study demonstrates the potential impact of device shape and size on valve mechanics. Further expanding our study to more valves may allow for universal recommendations in the future. [1] Nath, J. et al., JACC, 43:405-409, 2004. [2] Mathur, M. et. al., Eng with Comp, 38:3835-3848, 2022."
Investigating the Permissive Environment of Perisynaptic Astroglia for Information Storage in the Dentate Gyrus
(2022-09-29) Nam, A.J.; Kuwajima, M.; Mendenhall, J.M.; Hubbard, D.D.; Hanka, D.C.; Parker, P.H.; Wetzel, A.; Bartol, T.M.; Sejnowski, T.J.; Abraham, W.C.; Harris, K.M.
Perisynaptic astroglial processes (PAPs), are active modulators of neuronal activity and directly contribute to information processing in the brain. Both in vivo and in vitro experiments have demonstrated that PAPs undergo activity-dependent structural changes. Thus, here we employ cutting-edge resources at the Texas Advanced Computing Center (TACC) to explore PAP structural remodeling associated with long-term potentiation (LTP) and long-term depression (LTD) that may help support local changes in information processing. Long-term potentiation (LTP) and long-term depression (LTD), widely accepted cellular mechanisms of learning and memory, were induced in vivo in the awake adult rat hippocampal dentate gyrus. LTP induction in the middle molecular layer (MML) was achieved by delta-burst stimulation in the medial perforant pathway, a procedure that produced concurrent long-term depression (cLTD) in the outer molecular layer (OML). The contralateral control hemisphere received only baseline stimulation to the medial perforant path. Three-dimensional electron microscopy (3DEM) offers significant advantages over two-dimensional approaches including a more complete view of ultrastructure in all X-Y-Z planes. AlignEM Swift, the state-of-the-art interactive application available at TACC, is integral for achieving the standard of perfect serial section image alignment needed for 3DEM analysis. Furthermore, Blender at TACC, equipped with the computing power of TACC’s supercomputers, similarly facilitates large-scale and realistic PAP reconstructions for visualization and quantitative mesh analysis. Changes to PAP ultrastructure have important implications on the spatiotemporal dynamics of astrocyte calcium signaling. Thus, TACC resources will further enable computational modeling to investigate the functional consequences of PAP morphological changes. Preliminary analysis suggests that more than 80% of all dentate gyrus synapses exhibit some degree of PAP apposition at the axon-spine interface (ASI). Results from this study made possible using TACC systems will contribute to our overall understanding of the cellular mechanism of information processing and the role of specifically astrocytes in this process.
Investigating Vibrio Cholerae ToxT-Ligand Interactions Using GROMACS
(2023-09-28) Castellanos, Hugo Villar; Touhami, Ahmed; Hanke, Andreas
Vibrio cholerae is a bacterium responsible for the potentially fatal disease known as Cholera, and the ToxT protein is responsible for the transcription of most virulence genes in the bacterium. The purpose of this preliminary study is to investigate the binding affinity of a ligand that attaches to the DNA binding domain in ToxT using the GROMACS molecular dynamics package. Based on a literature review, a drug-like ligand (ZINC database id 14749003) was selected for this study. Docking of the ligand to ToxT using Molegro Virtual Docker (MVD) showed two prevalent binding sites (residues Arg214 and Arg187); the ligand binds to the protein via two hydrogen bonds with each site, with binding energies of -1.35 kcal/mol and -0.8 kcal/mol for Arg214, -1.6 kcal/mol and -1.46 kcal/mol for Arg187. Then analysis of a 1-ns molecular dynamics run (on Lonestar6) of the docked ligand-protein complex showed that the attachment of the ligand to the protein was relatively stable, with a maximum hydrogen bond occupancy rate of 31.7% and with the top donor-acceptor pairs involving residues Arg214 and Arg187, which validates the result from MVD. The conformation of the complex was also stable with a radius of gyration fluctuating around a value of 1.9 nm.
Job Losses, Marriage Troubles and Rich Uncles: Foreclosure Prevention Policy when Borrowers Hold Private Information about their Financial Health
(2022-09-29) Kytömaa, Lauri
My dissertation studies foreclosure prevention in environments where borrowers have an incentive to appear distressed in order to receive mortgage reductions. Such behavior is possible when borrowers have knowledge about their abilities to repay debt that cannot be observed by their lenders. Using a sample of Fannie Mae loans originated in California between 2004 and 2007, I show that mortgage providers only offer debt relief when they are highly informed about borrower default probabilities. I then use the estimated model to explore the effects of the Federal Home Affordability Modification Program, which was launched in 2009 in response to the Great Recession. I find that subsidies offered to banks under the program were more effective at preventing foreclosures in loans originated earlier in the 2000’s, even though banks tended to be equally well informed about borrower financial health in all sample cohorts. The results suggest that government subsides decreased foreclosures by 7.2% for loans originated in 2004, but that this rate steadily declines to 1.0% for loans originated in 2007. I also find that the average subsidy expenditure per prevented foreclosure increased from $17,000 to $150,000 between my sample origination cohorts. Jointly, the results offer a comprehensive look at how borrower financial well- being and the behavior of financial institutions influence debt relief policy. This project benefits greatly from access to TACC resources. Estimation uses a maximum likelihood routine in which I solve for a high-dimensional grid that rationalizes bank behavior, and then match model-predicted loan outcome probabilities to data. Solving for bank behavior is conducive to parallel computing since the optimum can be computed independently for every set of inputs. Leveraging many nodes allows me to solve for optimal policy at 60.2 million input combinations in under an hour. All numerical maximization takes place using Python on the Stampede2 cluster.
The Lewis Assisted C-CN Bond Activation of Benzonitrile Using Zerovalent Group 10 Metal Complexes with Dippe Ligand
(2023-09-28) Escobar, Roberto; Atesin, Abdurrahman; Jones, William D.; Ateşin, Tülay
The exploration of C-C bond activation is significantly propelled by homogeneous transition-metal catalysts, capturing attention for their role in industrial applications and intricate organic molecule synthesis. Nonetheless, the formidable challenge of effecting mild and homogeneous activation of thermodynamically stable and kinetically inert C—C σ-bonds persists. Presently, prevailing activation strategies predominantly entail systems driven by strain relief or aromatization, excepting the intriguing activation of unstrained C—CN bonds within nitriles. Prior investigations leveraging the [(dippe)Ni] fragment alongside Lewis acids have unveiled the manipulation of reaction mechanisms concerning the C-CN bond in benzonitrile. This is suggested in yielding faster reaction rates and nitrogen lone pair coordination that imparts stability through steric bulk or charge redistribution, thereby facilitating C-CN bond cleavage. By employing advanced computational tools, encompassing Gaussian16 and GaussView6 to perform density functional theory (DFT) calculations, this study delves into the C-CN bond activation of benzonitrile. This is achieved by replacing [(dippe)Ni](Ph)(CN) with heavier group 10 metals (Pd and Pt) in both reaction and product substrates. This strategic substitution enables optimizations and novel potential energy surface (PES) scans, guiding the search for intermediates and transition states, with structural validation through ChimeraX ensuring 3D accuracy and predicting the reaction pathway. Furthermore, the influence of Lewis acids (BPh3 and BF3) is examined, elucidating their effects on proposed reaction mechanisms, thermodynamics, and kinetics in comparison to those without Lewis acids. Additional refinements encompass gas phase and solvent corrections (THF and toluene), employing GoodVibes v3.2 for vibrational analysis and thermodynamic predictions, and addressing functional (B3YLP) limitation on π-system interaction via empirical dispersion corrections for heightened result accuracy. This amalgamation of transition-metal catalysts and Lewis acids establishes a foundation for innovative catalytic systems, poised to reshape chemical synthesis methodologies and chart transformative routes toward groundbreaking synthetic pathways.
Machine Learning Hub for Tapis
(2023-09-28) Indrakusuma, Dhanny; Freeman, Nathan; Stubbs, Joe
"Machine learning is indispensable for extracting insights from intricate datasets, expediting data analysis, and enabling cross-disciplinary decision-making. However, the complexity of machine learning models can hinder non-technical users, necessitating for user-friendly tools. Within the Cloud and Interactive Computing group (CIC) at the Texas Advanced Computing Center (TACC), we are actively developing a Machine Learning Hub (ML Hub) API for the Tapis Framework. Comprising accessible microservices, each with an independent REST API implemented in Python's Flask and codified with OpenAPI v3 definitions, our research aims to enhance the experiences of developers, scientists, and researchers. The integration of Hugging Face's API into ML Hub provides open-source pre-trained models for state-of-the-art AI capabilities. Currently, ML Hub's Models Overview and Models Download functions offer a gateway for non-technical users to explore and download machine learning models, authenticated using a JSON Web Token (JWT) from the Tapis Authenticator API. Future developments encompass implementing the Inference Client and Training Engine, seamlessly integrating with the Tapis UI in React and Typescript. Key features of ML Hub: 1. Models Overview: A portal showcasing top Hugging Face models with filtering options. 2. Models Download: Users can obtain specific models, with options to either download a binary file of the model or a zip file containing the model's repository, cached in a version-aware manner. 3. Inference Client: Facilitating server initiation for machine learning model inference on TACC's HPC cluster, enabling rapid prototyping. 4. Training Engine: Enabling users to fine-tune models and showcase them on TACC's HPC cluster, removing technical complexities. This research contributes to the broader discourse on democratizing machine learning's potential, by providing user- friendly access to state-of-the-art models and addressing non-technical users' challenges. We hope that this project will foster innovative collaboration and user engagement, paving the way for an inclusive and impactful future in machine learning research."
Medical Data Augmentation via ChatGPT: A Case Study on Medication Identification and Medication Event Classification
(2023-09-28) Sarker, Shouvon; Li, Xiangfang; Dong, Xishuang; Qian, Lijun
The identification of key factors such as medications, diseases, and relationships within electronic health records and clinical notes has a wide range of applications in the clinical field. In the N2C2 2022 competitions, various tasks were presented to promote the identification of key factors in electronic health records (EHRs) using the Contextualized Medication Event Dataset (CMED). Pretrained large language models (LLMs) demonstrated exceptional performance in these tasks. This study aims to explore the utilization of LLMs, specifically ChatGPT, for data augmentation to overcome the limited availability of annotated data for identifying the key factors in EHRs. Additionally, different pre-trained BERT models, initially trained on extensive datasets like Wikipedia and MIMIC, were employed to develop models for identifying these key variables in EHRs through fine-tuning on augmented datasets. The experimental results of two EHR analysis tasks, namely medication identification and medication event classification, indicate that data augmentation based on ChatGPT proves beneficial in improving performance for both medication identification and medication event classification.
Modeling Gas, Hydrates, and Slope Stability on the U.S. Atlantic Margin during Pleistocene Glaciations
(2023-09-28) Carty, Olin; Daigle, Hugh
Dissociation of methane hydrates in shallow marine sediments due to increasing global temperatures can lead to the venting of methane gas or seafloor destabilization. Along the U.S. Atlantic margin there is a well-documented history of slope failure and numerous gas seeps have been recorded. Several studies have linked slope failure to gas seepage and hydrate dissociation driven by glacial-interglacial transitions, but this linkage has not been quantitatively demonstrated. Along the shelf edge, in an area where shallow methane gas seeps have been identified, we modeled methane gas and hydrate formation over the last 120,000 to simulate a glacial-interglacial cycle. The development of hydrate and gas during this time was modeled using the PFLOTRAN software from Sandia National Laboratories, a parallel subsurface flow code. At 100-year intervals during this simulation, we calculated the factor of safety throughout the modeled sediment column. Factor of safety compares the shearing and resisting stresses of a slope and can be used to determine if sediment failure is likely to occur in an area. Modeling seafloor depths between 200-1000 m we predicted gas and hydrate development and calculated the associated factor of safety over time to determine if sediment failure was likely to be caused by hydrate dissociation. Parallelizing this code, we used Lonestar6 to run the one-dimensional fluid flow model and factor of safety model at 16044 individual locations in the region between 29°N – 45°N and 82°W – 66°W at a resolution of 1 x 1 arcminutes. We found that hydrate dissociation alone is unlikely to cause sediment failure in the region, implying that an additional driving force would be necessary for failure to occur. In addition, we see a shift down slope of when the minimum factor of safety is likely to occur and the depth below seafloor at which this minimum occurs.
Modeling of Tunable Spike-Timing-Dependent Plasticity (STDP) Neuromorphic Devices using Magnetic Skyrmion Manipulation Chambers
(2023-09-28) Khodzhaev, Zulfidin; Incorvia, Jean Anne
"Magnetic skyrmions are promising candidates for implementing synaptic plasticity in neuromorphic computing due to their dynamic and nonvolatile nature [1]. This work presents the modeling of a neuromorphic device with three magnetic skyrmion manipulation chambers, Fig. 1, to emulate spike-timing-dependent plasticity (STDP) [2]. The middle chamber stores the synaptic weight as the skyrmion count, while the side chambers generate pre- and post-synaptic spikes. By controlling the timing and magnitude of currents applied to the chambers, the final skyrmion count can be tuned to demonstrate STDP learning rules [3]. The device exhibits configurable rates of weight update, emulating location-dependent plasticity along dendrites. The modeled skyrmion device demonstrates adaptability in implementing diverse synaptic plasticity functions for efficient neuromorphic computing. [1] A. Fert, N. Reyren, and V. Cros, Nature Reviews Materials 2017 2:7 2(7), 1–15 (2017). [2] G.Q. Bi, and M.M. Poo, Journal of Neuroscience 18(24), 10464–10472 (1998). [3] R.C. Froemke, M.M. Poo, and Y. Dan, Nature 2005 434:7030 434(7030), 221–225 (2005)."
Multi-GPU FFT Matvec for Inverse Problems Involving Shift-Invariant Systems
(2023-09-28) Venkat, Sreeram; Fernando, Milinda; Henneking, Stefan; Ghattas, Omar
Hessian-based algorithms for the solution to inverse problems typically require many actions of the Hessian matrix on a vector (matvecs). A direct approach is often computationally intractable for problems with high-dimensional parameter fields or expensive-to-evaluate forward models. For systems that exhibit shift-invariance (e.g. autonomous systems) structure in their discretized form, the discretized linear parameter-to-observable (p2o) maps are block Toeplitz matrices. Moreover, considering causality for time-invariant systems the p2o map and its adjoint are lower- and upper-triangular block Toeplitz, respectively. By exploiting this structure, Hessian matrices for these types of systems can be compactly represented and Hessian matvecs can be efficiently computed through scalable multi-GPU FFT matvecs. Compact representation follows directly from the definition of Block Toeplitz matrices. Fast matrix-vector multiplication is achieved by embedding the block Toeplitz matrix within a block circulant matrix which is diagonalized by the Discrete Fourier Transform. The matrix-vector product then becomes an element-wise vector operation in Fourier space. Furthermore, the action of the adjoint p2o map corresponds to simply applying the complex conjugate in Fourier space, eliminating the need to separately store the Fourier-transformed forward and adjoint map. Exploiting the triangular block Toeplitz structure in this way yields memory savings proportional to the number of time steps Nt and a computational speedup of O(Nt/ log(Nt)). In the context of explicit methods that are suitable for GPU-based computation, the number of time steps is typically very large due to the CFL condition, making the savings of the algorithm substantial. We develop a multi-GPU FFT matvec code for Hessians corresponding to block Toeplitz p2o matrices utilizing the cuFFT and NCCL libraries. Our implementation achieves 75-90% of the maximum memory bandwidth on NVIDIA A100 80GB GPUs for all custom GPU kernels — which correspond to memory-bound operations. We also show strong and weak multi-GPU scaling on the Frontera RTX nodes with up to 81 GPUs.