Search NASA⌕ Search

SEARCH · Search NASA

Results for “Machine Learning in Network Science”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 1,045 records · Page 58

Cooperative Research and Development Agreement With Georgetown University Report: National Institutes of Health, National Center for Advancing Translational Sciences Clinical and Translational Science Award

Oak Ridge National Laboratory (ORNL) is participating with Georgetown University (GU) as a subrecipient in response to the National Institutes of Health (NIH), National Center for Advancing Translational Sciences (NCATS) Clinical and Translational Science Award (CTSA) (U54 Clinical Trial Optional) funding opportunity announcement. This Cooperative Research and Development Agreement is put in to place to facilitate the development and implementation of clinical interventions that demonstrably improve human health is currently a complex, recursive, and inefficient process that leads to delays of years or decades before discoveries in biomedical research result in health benefits for patients and communities. NCATS conducts and supports research in the science of translation, to discover the mechanistic and operational principles of the intervention development and dissemination process, thereby providing the scientific foundation for improvements in translational efficiency that will accelerate the realization of interventions that improve human health. Under NCATS’ leadership, the CTSA Program supports a national network of medical research institutions called hubs. GU is the lead institution in one of the NIH hubs that was created as a result of a previous NIH CTSA. The missions of the GU have historically included the advancement of health through research in the clinical and biomedical sciences, the education of future leaders in medical and nursing practice and academia, and the provision of compassionate and scientifically competent patient care and service to the Washington, DC community and the nation. GU is the lead institution for the Georgetown-Howard Universities Center for Clinical and Translational Science (GHUCCTS), a multi-institutional partnership of medical research institutions forged from a desire to promote clinical research and translational science. Through multiple collaborations among these institutions, GHUCCTS is transforming clinical research and translational science in order to bring new scientific advances to health care. Oak Ridge National Laboratory is the Department of Energy's (DOE) largest science and energy laboratory. Managed since April 2000 by a partnership of the University of Tennessee and Battelle, ORNL was established in 1943 as a part of the secret Manhattan Project to pioneer a method for producing and separating plutonium. During the 1950s and 1960s, ORNL became an international center for the study of nuclear energy and related research in the physical and life sciences. With the creation of DOE in the 1970s, ORNL's mission broadened to include a variety of energy technologies and strategies. Today the laboratory supports the nation with a peacetime science and technology mission that is just as important as, but very different from, its role during the Manhattan Project. ORNL is home to the world's premier center for high performance supercomputing to enable scientific discovery. ORNL has extensive expertise in various areas of computer science that are uniquely situated to support GU. Additionally, ORNL’s leading computational user facilities present a unique opportunity to leverage the largest scale machines for open science in support of the stated mission of the NCATS CTSA. ORNL's partnership with GU will offer unparalleled opportunity in data analytics, deep-learning, artificial intelligence, and urban dynamics.

59 BASIC BIOLOGICAL SCIENCES↗

Machine learning pipeline for denoising low signal-to-noise ratio and out-of-distribution transmission electron microscopy datasets

High-resolution transmission electron microscopy (HRTEM) is crucial for observing material’s structural and morphological evolution at Angstrom scales, but the electron beam can alter these processes. Devices such as CMOS-based direct-electron detectors operating in electron-counting mode can be utilized to substantially reduce the electron dosage. However, the resulting images often lead to a low signal-to-noise ratio, which requires frame integration that sacrifices temporal resolution. Several machine learning (ML) models have been recently developed to successfully denoise HRTEM images. Yet, these models are often computationally expensive, and their inference speeds on GPUs are outpaced by the imaging speed of advanced detectors, precluding in situ analysis. Furthermore, the performance of these denoising models on datasets with imaging conditions that deviate from the training datasets has not been evaluated. To mitigate these gaps, we propose a new self-supervised ML denoising pipeline specifically designed for time-series HRTEM images. This pipeline integrates a blind-spot convolution neural network with pre-processing and post-processing steps, including drift correction and low-pass filtering. Results demonstrate that our model outperforms various other ML and non-ML denoising methods in noise reduction and contrast enhancement, leading to improved visual clarity of atomic features. Additionally, the model is drastically faster than U-Net-based ML models and demonstrates excellent out-of-distribution generalization. The model’s computational inference speed is in the order of milliseconds per image, rendering it suitable for application in in-situ HRTEM experiments.

36 MATERIALS SCIENCE↗

Argonne Leadership Computing Facility: 2021 Operational Assessment Report

This Operational Assessment Report describes how the Argonne Leadership Computing Facility (ALCF) met or exceeded every one of its goals for calendar year (CY) 2021 as an advanced scientific computing center. In CY 2021, the ALCF operated its production resource, Theta, an Intel-based Cray XC40 system (11.7-petaflops) augmented with 24 NVIDIA DGX A100-based nodes (3.9-petaflops) that supports diverse workloads, integrating data analytics with artificial intelligence (AI) training and learning in a single platform. In 2021, we began deploying Polaris, our newest 40- petaflops system, and augmented this powerful testbed system with an additional 28 nodes to support the integration of real-time experiments and HPC resources. We also deployed our two largest storage systems yet, named Grand and Eagle, that will bring new services to our users and will power data-driven research for years to come. Last year, Theta delivered a total of 20.8 million node-hours to 16 Innovative and Novel Computational Impact on Theory and Experiment (INCITE) projects and 7.2 million node-hours to ASCR Leadership Computing Challenge (ALCC) projects (32 awarded during the 2020–2021 ALCC year and 17 awarded during the 2021–2022 ALCC year), as well as substantial support to Director’s Discretionary (DD) projects (5.5 million node-hours). As Table ES.1 shows, Theta performed exceptionally well in terms of overall availability (95.1 percent), scheduled availability (99.4 percent), and utilization (98.1 percent; Table 2.1). As of the submission date of this document, ALCF’s user community has published 249 papers in high-quality, peer-reviewed journals and technical proceedings. At the 2021 International Conference for High Performance Computing, Networking, Storage and Analysis (SC’21), Argonne researchers won two HPCwire Readers’ Choice Awards and were part of a Gordon Bell Prize finalist team recognized for developing an AI-enabled, multi-resolution simulation framework for studying complex biomolecular machines. Their framework was used to observe the SARS-CoV-2 replication-transcription machinery in action, by directly integrating experimental data. ALCF also provided a comprehensive program of high-performance computing (HPC) support services to help our community make productive use of the facility’s diverse and growing collection of resources. We are now entering the exascale era, with exascale machines being planned for national laboratories across the country, including Aurora at Argonne National Laboratory (Argonne) in 2023. ALCF researchers have been leading and guiding numerous strategic activities that will push the boundaries of what’s possible in computational science and engineering and allow us to deliver science on day one.

97 MATHEMATICS AND COMPUTING↗

Super-Resolution from Space: Using MERRA-2 and MAIAC Satellite Imagery to Produce Daily Continuous 1 km PM2.5 Estimates

PM2.5 measurements from ground stations are the gold standard when available, but the expense and coverage of such stations limits widespread monitoring. Having accurate PM2.5 estimates outside the range of these stations is important for monitoring this crucial aspect of air quality. The goal of this project is to produce daily 1 km continuous PM2.5 estimates for the contiguous US relying primarily on satellite-derived data sources. This is important because models based on such data can be more easily expanded outside the study area and produce global estimates as well. The temporal availability of such data products is often weekly/daily, unlike land-use products with are often available at a yearly or worse temporal resolution. To achieve our goal, we use a couple of different deep neural network architectures to produce PM2.5 measurements at 10 km and 1 km resolution. We use two model architectures, a UNET-like model and a GAN-based model. We train both models using MERRA-2 data and MAIAC AOD data scaled to 10 km and 1 km for the two different prediction resolutions. MERRA-2 imagery is data rich with a wide range of geospatial variables at 50 km and has long historical availability (beginning in 1980). We’re also using higher spatial resolution MAIAC data at 1 km to provide finer resolution spatial context. This essentially leverages the spatial resolution of MAIAC data and the “wider” information of MERRA-2 data to predict PM2.5. For the target data we’re using a modeled 1 km PM2.5 dataset produced by Harvard to pre-train our models and then fine-tune our models using ground station measurements. Not only are our results comparable with the performance of the Harvard dataset, but can be generalized to any area or time where MERRA-2 and MAIAC data is available.

satellite imagery↗

Machine learning elastic constants of multi-component alloys

The present manuscript explores application of machine learning methods for determining elastic constants and other derived mechanical properties of multi-component alloys. Here, a number of machine learning models, including linear regression, neural network and random forest based models, are trained and tested on a dataset of binary alloys generated using density functional theory (DFT) calculations and spanning over a large number of elemental species in the periodic table. Starting with a wide range of simple and easily accessible compositionally-averaged elemental features, a correlation-based feature selection strategy was used to systematically down-select a set of most relevant features towards the prediction of the elasticity tensor components. The true predictive performance and the associated uncertainties of the models were established by testing on unseen data and bootstrapping, respectively. A single and pair-wise feature partial dependence analysis was performed to visualize the average property trends in the multi-dimensional feature space in order to further understand the achieved predictive performance. The utility of the trained model is further demonstrated by obtaining sufficiently accurate yet highly efficient approximations for bulk modulus, Young’s modulus, shear modulus and Poisson’s ratio for alloys beyond the binary space (i.e., two-component alloys) on which the model was originally trained. More importantly, we test and validate the predictive performance of the developed model directly against the experimentally measured elastic constants of technologically relevant multi-component alloys (such as, Ni- and Ti-based alloys). Finally, utility of such a data-enabled route is demonstrated by predicting the possible range of various elastic properties for vast composition space available within the five component Ni-Cr-Fe-Mo-W alloy system in a high-throughput manner.

36 MATERIALS SCIENCE↗

Track Seeding and Labelling with Embedded-space Graph Neural Networks

To address the unprecedented scale of HL-LHC data, the Exa.TrkX project is investigating a variety of machine learning approaches to particle track reconstruction. The most promising of these solutions, graph neural networks (GNN), process the event as a graph that connects track measurements (detector hits corresponding to nodes) with candidate line segments between the hits (corresponding to edges). Detector information can be associated with nodes and edges, enabling a GNN to propagate the embedded parameters around the graph and predict node-, edge- and graph-level observables. Previously, message-passing GNNs have shown success in predicting doublet likelihood, and we here report updates on the state-of-the-art architectures for this task. In addition, the Exa.TrkX project has investigated innovations in both graph construction, and embedded representations, in an effort to achieve fully learned end-to-end track finding. Hence, we present a suite of extensions to the original model, with encouraging results for hitgraph classification. In addition, we explore increased performance by constructing graphs from learned representations which contain non-linear metric structure, allowing for efficient clustering and neighborhood queries of data points. We demonstrate how this framework fits in with both traditional clustering pipelines, and GNN approaches. The embedded graphs feed into high-accuracy doublet and triplet classifiers, or can be used as an end-to-end track classifier by clustering in an embedded space. A set of post-processing methods improve performance with knowledge of the detector physics. Finally, we present numerical results on the TrackML particle tracking challenge dataset, where our framework shows favorable results in both seeding and track finding.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗

Prediction of chronic kidney disease progression using recurrent neural network and electronic health records

Chronic kidney disease (CKD) is a progressive loss in kidney function. Early detection of patients who will progress to late-stage CKD is of paramount importance for patient care. To address this, we develop a pipeline to process longitudinal electronic heath records (EHRs) and construct recurrent neural network (RNN) models to predict CKD progression from stages II/III to stages IV/V. The RNN model generates predictions based on time-series records of patients, including repeated lab tests and other clinical variables. Our investigation reveals that using a single variable, the recorded estimated glomerular filtration rate (eGFR) over time, the RNN model achieves an average area under the receiver operating characteristic curve (AUROC) of 0.957 for predicting future CKD progression. When additional clinical variables, such as demographics, vital information, lab test results, and health behaviors, are incorporated, the average AUROC increases to 0.967. In both scenarios, the standard deviation of the AUROC across cross-validation trials is less than 0.01, indicating a stable and high prediction accuracy. Our analysis results demonstrate the proposed RNN model outperforms existing standard approaches, including static and dynamic Cox proportional hazards models, random forest, and LightGBM. The utilization of the RNN model and the time-series data of previous eGFR measurements underscores its potential as a straightforward and effective tool for assessing the clinical risk of CKD patients concerning their disease progression.

60 APPLIED LIFE SCIENCES↗

Liquid–Vapor Phase Equilibrium in Molten Aluminum Chloride (AlCl 3 ) Enabled by Machine Learning Interatomic Potentials

Molten salts are promising candidates in numerous clean energy applications, where knowledge of thermophysical properties and vapor pressure across their operating temperature ranges is critical for safe operations. Due to challenges in evaluating these properties using experimental methods, fast and scalable molecular simulations are essential to complement the experimental data. In this study, we developed machine learning interatomic potentials (MLIP) to study the AlCl 3 molten salt across varied thermodynamic conditions (T = 473–613 K and P = 2.7–23.4 bar), which allowed us to predict temperature-surface tension correlations and liquid–vapor phase diagram from direct simulations of two-phase coexistence in this molten salt. Two MLIP architectures, a Kernel-based potential and neural network interatomic potential (NNIP), were considered to benchmark their performance for AlCl 3 molten salt using experimental structure and density values. The NNIP potential employed in two-phase equilibrium simulations yields the critical temperature and critical density of AlCl 3 that are within 10 K (∼3%) and 0.03 g/cm 3 (∼7%) of the reported experimental values. An accurate correlation between temperature and viscosities is obtained as well. In doing so, we report that the inclusion of low-density configurations in their training is critical to more accurately represent the AlCl 3 system across a wide phase-space. The MLIP trained using PBE-D3 functional in the ab initio molecular dynamics (AIMD) simulations (120 atoms) also showed close agreement with experimentally determined molten salt structure comprising Al 2 Cl 6 dimers, as validated using Raman spectra and neutron structure factor. Furthermore, the PBE-D3 as well as its trained MLIP showed better liquid density and temperature correlation for AlCl 3 system when compared to several other density functionals explored in this work. Overall, the demonstrated approach to predict temperature correlations for liquid and vapor densities in this study can be employed to screen nuclear reactors-relevant compositions, helping to mitigate safety concerns.

Ab initio molecular dynamics↗

Model Evaluation and Intercomparison of Marine Warm Low Cloud Fractions With Neural Network Ensembles

Abstract Low cloud fractions (LCFs) and meteorological factors (MFs) over an oceanic region containing multiple cloud regimes are examined for three data sets: one Energy Exascale Earth System Model (E3SM) simulation with the default 72‐layer vertical grid (E3SM72), another one with 8‐times vertical resolution via the Framework for Improvement by Vertical Enhancement (E3SM 8), and one with MFs from ERA5 reanalysis and LCFs from the CERES SSF product (ERA5‐SSF). Neural networks (NNs) are trained to capture the relationship between MFs and LCF and to select the best‐performing MF subsets for predicting LCF. NN ensembles are used to (a) confirm the performance of selected MF subsets, (b) to serve as proxy models for each data set to predict LCFs for MFs from all data sets, and (c) to classify MFs into those in shared and uniquely occupied MF subspaces. Overall, E3SM72 and E3SM 8 have large fractions of MFs in shared MF subspace, but less so near the Californian and Peruvian stratocumulus decks. E3SM 8 and ERA5 have small fractions of MFs in shared MF subspace but greater than E3SM72 and ERA5, especially in the Southeast Pacific. The differences in LCFs between three pairs of data sets are decomposed into those associated with the differences in the LCF‐MF relationship and those involving different MFs. Given the same MFs, LCFs produced by E3SM 8 are greater than those produced by E3SM72 but are still different from those in ERA5‐SSF. In general, the shift in MFs dominates the difference in the LCFs.

54 ENVIRONMENTAL SCIENCES↗

NLML: A Deep Neural Network Emulator for the Exact Nonlinear Interactions in a Wind Wave Model

Nonlinear wave interactions describe the resonant energy transfer between wave components, playing a fundamental role in the evolution of ocean wave spectra. Nonlinear wave interactions significantly influence wave growth and development, making them essential for accurate wave modeling. However, resolving the full six-dimensional Boltzmann integral of the exact nonlinear wave interactions (Webb-Resio-Tracy method, WRT) is computationally expensive, limiting its application in real-time operational wave forecasting and for research purposes. Current approximations, such as the Discrete Interaction Approximation (DIA), prioritize computational speed over accuracy, resulting in significant errors in wave mean parameters. Here, we introduce NLML, a machine learning (ML) emulator designed to approximate the exact nonlinear wave interactions within WAVEWATCH III (WW3), with the goal of achieving the accuracy of WRT while maintaining the stability and computational speed of DIA. By leveraging GPU capabilities such as half precision inference, we achieved substantial speedups, up to 136x mathematical equation faster than the WRT and only a modest 1.04x mathematical equation slowdown relative to DIA, while achieving 2x mathematical equation the accuracy of DIA in global wave spectral energy and mean wave parameters, with up to 7x mathematical equation higher accuracy in some regions. Unlike previous ML approaches, NLML maintained inherent stability throughout model integration in a standalone, year-long WW3 simulation, without requiring additional constraints. Our new ML parameterization bridges the gap between accuracy and efficiency, offering a promising alternative for improving wave modeling in operational settings and research purposes.

16 TIDAL AND WAVE POWER↗

Development of Short-Term Forecasting Models Using Plant Asset Data and Feature Selection

Nuclear power plants collect and store large volumes of heterogeneous data from various components and systems. With recent advances in machine learning (ML) techniques, these data can be leveraged to develop diagnostic and short-term forecasting models to better predict future equipment condition. Maintenance operations can then be planned in advance whenever degraded performance is predicted, thus resulting in fewer unplanned outages and the optimization of maintenance activities. This enables lower maintenance costs and improves the overall economics of nuclear power. This paper focuses on developing a short-term forecasting process that leverages a feature selection process to distill large volumes of heterogeneous data and predict specific equipment parameters. A variety of feature selection methods, including Shapley Additive Explanations (SHAP) and variance inflation factor (VIF), were used to select the optimal features as inputs for three ML methods: long short-term memory (LSTM) networks, support vector regression (SVR), and random forest (RF). Each combination of model and input features was used to predict a pump bearing temperature both 1 and 24 hours in advance, based on actual plant system data. The optimal inputs for the LSTM and SVR were selected using the SHAP values, while the optimal input for the RF consisted solely of the response variable itself. Each model produced similar 1-hour-ahead predictions, with root mean square errors (RMSEs) of roughly 0.006. For the 24-hour-ahead predictions, differences could be seen between LSTM, SVR, and RF, as reflected by model performances of 0.036 +- 0.014, 0.0026 +- 0, and 0.063 +- 0.004 RMSE, respectively. As big data and continuous online monitoring become more widely available, the proposed feature selection process can be used for many applications beyond the prediction of process parameters within nuclear infrastructure.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗

Scaling kinetic Monte-Carlo simulations of grain growth with combined convolutional and graph neural networks

Graph neural networks (GNN) have emerged as a promising machine learning method for microstructure simulations such as grain growth. However, accurate modeling of realistic grain boundary networks requires large simulation cells, which GNN has difficulty scaling up to. To alleviate the computational costs and memory footprint of GNN, we suggest a hybrid architecture combining a convolutional neural network (CNN) based bijective autoencoder to compress the spatial dimensions, and a GNN that evolves the microstructure in the latent space of reduced spatial sizes. Our results demonstrate that the new design significantly reduces computational costs with using fewer message passing layer (from 12 down to 3) compared with GNN alone. The reduction in computational cost becomes more pronounced as the spatial size increases, indicating strong computational scalability. For the largest mesh evaluated (160 3 ), our method reduces memory usage and runtime in inference by 117× and 115×, respectively, compared with GNN-only baseline. More importantly, it shows higher accuracy and stronger spatiotemporal capability than the GNN-only baseline, especially in long-term testing. Such combination of scalability and accuracy is essential for simulating realistic material microstructures over extended time scales. The improvements can be attributed to the bijective autoencoder’s ability to compress information losslessly from spatial domain into a high dimensional feature space, thereby producing more expressive latent features for the GNN to learn from, while also contributing its own spatiotemporal modeling capability. Training data are generated from stochastic grain growth simulations, providing realistic variability for learning robust microstructure evolution. Comprehensive system validation confirms that the model is accurate, robust, and scalable.

36 MATERIALS SCIENCE↗

WE-H-BRC-06: A Unified Machine-Learning Based Probabilistic Model for Automated Anomaly Detection in the Treatment Plan Data

Purpose: The purpose of this work was to investigate the ability of a machine-learning based probabilistic approach to detect radiotherapy treatment plan anomalies given initial disease classes information. Methods In total we obtained 1112 unique treatment plans with five plan parameters and disease information from a Mosaiq treatment management system database for use in the study. The plan parameters include prescription dose, fractions, fields, modality and techniques. The disease information includes disease site, and T, M and N disease stages. A Bayesian network method was employed to model the probabilistic relationships between tumor disease information, plan parameters and an anomaly flag. A Bayesian learning method with Dirichlet prior was useed to learn the joint probabilities between dependent variables in error-free plan data and data with artificially induced anomalies. In the study, we randomly sampled data with anomaly in a specified anomaly space.We tested the approach with three groups of plan anomalies – improper concurrence of values of all five plan parameters and values of any two out of five parameters, and all single plan parameter value anomalies. Totally, 16 types of plan anomalies were covered by the study. For each type, we trained an individual Bayesian network. Results: We found that the true positive rate (recall) and positive predictive value (precision) to detect concurrence anomalies of five plan parameters in new patient cases were 94.45±0.26% and 93.76±0.39% respectively. To detect other 15 types of plan anomalies, the average recall and precision were 93.61±2.57% and 93.78±3.54% respectively. The computation time to detect the plan anomaly of each type in a new plan is ∼0.08 seconds. Conclusion: The proposed method for treatment plan anomaly detection was found effective in the initial tests. The results suggest that this type of models could be applied to develop plan anomaly detection tools to assist manual and automated plan checks. The senior author received research grants from ViewRay Inc. and Varian Medical System.

60 APPLIED LIFE SCIENCES↗

SMC 2021 : Analyzing Resource Utilization and User Behavior on Titan Supercomputer

Resource utilization statistics of submitted jobs on a supercomputer can help us understand how users from various scientific domains use HPC platforms and better design a job scheduler. We explore to generate insight regarding workload distribution and usage pattern domains from job scheduler trace, GPU failure information, and project-specific information collected from Titan supercomputer. Furthermore, we want to know how the scheduler performance varies over time and how the users' scheduling behavior changes following a system failure. These observations have the potential to provide valuable insight, which is helpful to prepare for system failures. These practices will help us develop and apply novel machine learning algorithms in understanding system behavior, requirement, and better scheduling of HPC systems. There are two datasets, RUR and GPU. RUR: This dataset is the job scheduler traces collected from the Titan supercomputerfrom 01/01/2015 to 07/31/2019 (2015.csv - 2019.csv). These were collected usingResource Utilization Report (RUR), a Cray-developed resource-usage data collectionand reporting system. It contains the usage information of its critical resources (CPU,Memory, GPU, and I/O) of each running job on Titan during that period [2]. ProjectAreas: Every job is associated with a project ID. TheProjectAreas.csvdatasetprovides a mapping of the project ID to its domain science. GPU: There have been some hardware-related issues in the GPUs in Titan that caused some GPUs to fail, sometimes irrecoverably during some job runs. This dataset provides information regarding these failures during the execution of the submitted jobs. GPUs on Titan are uniquely identified by a serial number (SN), and they are installed in a location. A GPU can be installed in a location, then removed from that location following a failure, and then re-installed in a different location after fixing the problem. If the failure can't be recovered, the GPU might be removed entirely from Titan. There are two prominent types of failures that resulted in the removal of GPUs from Titan: Double Bit Error (DBE) and Out of the Bus (OTB). The dataset (gc_full.csv) has the following fields: 1. SN : Serial number of a GPU 2. location : The location where it is installed 3. insert : The time when it was inserted into that location 4. remove : The time when it was removed from that location 5. duration : Amount of time the GPU spent in this location 6. out : If the device was taken out entirely w/o a re-installment into a new location. 7. event : If the GPU was taken out entirely, the reason for its removal. To learn more about this dataset, please refer to the git repositoryhttps://github.com/olcf/TitanGPULifeand the related publication [1]. References [1] George Ostrouchov, Don Maxwell, Rizwan A Ashraf, Christian Engelmann, MallikarjunShankar, and James H Rogers. Gpu lifetimes on titan supercomputer: Survival analysisand reliability. InSC20: International Conference for High Performance Computing,Networking, Storage and Analysis, pages 1-14. IEEE, 2020. [2] Feiyi Wang, Sarp Oral, Satyabrata Sen, and Neena Imam. Learning from five-yearresource-utilization data of titan system. In2019 IEEE International Conference onCluster Computing (CLUSTER), pages 1-6. IEEE, 2019.

42 ENGINEERING↗

Ultrafast processing of pixel detector data with machine learning frameworks

Modern photon science performed at high repetition rate free-electron laser (FEL) facilities and beyond relies on 2D pixel detectors operating at increasing frequencies (towards 100 kHz at LCLS-II) and producing rapidly increasing amounts of data (towards TB/s). This data must be rapidly stored for offline analysis and summarized in real time for online feedback to the scientists. While at LCLS all raw data has been stored, at LCLS-II this would lead to a prohibitive cost; instead, enabling real time processing of pixel detector data (dark, gain, common mode, background, charge summing, subpixel position, photon counting, data summarization) allows reducing the size and cost of online processing, offline processing and storage by orders of magnitude while preserving full photon information. This could be achieved by taking advantage of the compressibility of sparse data typical for LCLS-II applications. Faced with a similar big data challenge a decade ago, computer vision stimulated revolutionary advances in machine learning hardware and software. We investigated if these developments are useful in the field of data processing for high speed pixel detectors and found that typical deep learning models and autoencoder architectures failed to yield useful noise reduction while preserving full photon information, presumably because of the very different statistics and feature sets in computer vision and radiation imaging. However, the raw performance of modern frameworks like Tensorflow inspired us to redesign in Tensorflow mathematically equivalent versions of the state-of-the-art, “classical” algorithms used at LCLS. The novel Tensorflow models resulted in elegant, compact and hardware agnostic code, gaining 1 to 2 orders of magnitude faster processing on an inexpensive consumer GPU, reducing by 3 orders of magnitude the projected cost of online analysis and compression without photon loss at LCLS-II. The novel Tensorflow models also enabled ongoing development of a pipelined hardware system expected to yield an additional 3 to 4 orders of magnitude speedup, necessary for meeting the data acquisition and storage requirements at LCLS-II, potentially enabling acquiring every single FEL pulse at full speed. Computer vision a decade ago was dominated by hand-crafted filters; their structure inspired the deep learning revolution resulting in modern deep convolutional networks; similarly, our novel Tensorflow filters provide inspiration for designing future deep learning models for ultrafast and efficient processing and classification of pixel detector images at FEL facilities.

47 OTHER INSTRUMENTATION↗

Simultaneously improving accuracy and computational cost under parametric constraints in materials property prediction tasks

Abstract Modern data mining techniques using machine learning (ML) and deep learning (DL) algorithms have been shown to excel in the regression-based task of materials property prediction using various materials representations. In an attempt to improve the predictive performance of the deep neural network model, researchers have tried to add more layers as well as develop new architectural components to create sophisticated and deep neural network models that can aid in the training process and improve the predictive ability of the final model. However, usually, these modifications require a lot of computational resources, thereby further increasing the already large model training time, which is often not feasible, thereby limiting usage for most researchers. In this paper, we study and propose a deep neural network framework for regression-based problems comprising of fully connected layers that can work with any numerical vector-based materials representations as model input. We present a novel deep regression neural network, iBRNet, with branched skip connections and multiple schedulers, which can reduce the number of parameters used to construct the model, improve the accuracy, and decrease the training time of the predictive model. We perform the model training using composition-based numerical vectors representing the elemental fractions of the respective materials and compare their performance against other traditional ML and several known DL architectures. Using multiple datasets with varying data sizes for training and testing, We show that the proposed iBRNet models outperform the state-of-the-art ML and DL models for all data sizes. We also show that the branched structure and usage of multiple schedulers lead to fewer parameters and faster model training time with better convergence than other neural networks. Scientific contribution: The combination of multiple callback functions in deep neural networks minimizes training time and maximizes accuracy in a controlled computational environment with parametric constraints for the task of materials property prediction.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗

Scaling Deep Learning on GPU and Knights Landing clusters

The speed of deep neural networks training has become a big bottleneck of deep learning research and development. For example, training GoogleNet by ImageNet dataset on one Nvidia K20 GPU needs 21 days. To speed up the training process, the current deep learning systems heavily rely on the hardware accelerators. However, these accelerators have limited on-chip memory compared with CPUs. To handle large datasets, they need to fetch data from either CPU memory or remote processors. We use both self-hosted Intel Knights Landing (KNL) clusters and multi-GPU clusters as our target platforms. From an algorithm aspect, current distributed machine learning systems are mainly designed for cloud systems. These methods are asynchronous because of the slow network and high fault-tolerance requirement on cloud systems. We focus on Elastic Averaging SGD (EASGD) to design algorithms for HPC clusters. Original EASGD used round-robin method for communication and updating. The communication is ordered by the machine rank ID, which is inefficient on HPC clusters. First, we redesign four efficient algorithms for HPC systems to improve EASGD's poor scaling on clusters. Async EASGD, Async MEASGD, and Hogwild EASGD are faster \textcolor{black}{than} their existing counterparts (Async SGD, Async MSGD, and Hogwild SGD, resp.) in all the comparisons. Finally, we design Sync EASGD, which ties for the best performance among all the methods while being deterministic. In addition to the algorithmic improvements, we use some system-algorithm codesign techniques to scale up the algorithms. By reducing the percentage of communication from 87% to 14%, our Sync EASGD achieves 5.3x speedup over original EASGD on the same platform. We get 91.5% weak scaling efficiency on 4253 KNL cores, which is higher than the state-of-the-art implementation.

60 APPLIED LIFE SCIENCES↗

Using Machine Learning to Generate a GISS ModelE Calibrated Physics Ensemble (CPE)

A neural network (NN) surrogate of the NASA GISS ModelE atmosphere (version E3) is trained on a perturbed parameter ensemble (PPE) spanning 45 physics parameters and 36 outputs. The NN is leveraged in a Markov Chain Monte Carlo (MCMC) Bayesian parameter inference framework to generate a second posterior constrained ensemble coined a “calibrated physics ensemble,” or CPE. The CPE members are characterized by diverse parameter combinations and are, by definition, close to top-of-atmosphere radiative balance, and must broadly agree with numerous hydrologic, energy cycle and radiative forcing metrics simultaneously. Global observations of numerous cloud, environment, and radiation properties (provided by global satellite products) are crucial for CPE generation. The inference framework explicitly accounts for discrepancies (or biases) in satellite products during CPE generation. We demonstrate that product discrepancies strongly impact calibration of important model parameter settings (e.g., convective plume entrainment rates; fall speed for cloud ice). Structural improvements new to E3 are retained across CPE members (e.g., stratocumulus simulation). Notably, the framework improved the simulation of shallow cumulus and Amazon rainfall while not degrading radiation fields, an upgrade that neither default parameters nor Latin Hypercube parameter searching achieved. Analyses of the initial PPE suggested several parameters were unimportant for output variation. However, many “unimportant” parameters were needed for CPE generation, a result that brings to the forefront how parameter importance should be determined in PPEs. From the CPE, two diverse 45-dimensional parameter configurations are retained to generate radiatively-balanced, auto-tuned atmospheres that were used in two E3 submissions to CMIP6.

54 ENVIRONMENTAL SCIENCES↗