The use of data produced by scientific instruments, such as the Advanced Photon Source Upgrade (APS-U), to train and fine-tune machine learning models is becoming increasingly challenging due to high data production rates, large data volumes, and the growing complexity of machine learning models. To address these challenges, researchers have developed frameworks like fairDMS to efficiently organize vast amounts of data and models for rapid querying when model degradation is detected. However, the complexity of these frameworks and the physically distributed nature of experimental facilities complicate their deployment.Here we introduce a high-performance model and data management framework for machine learning, M2ML. In contrast to previous frameworks, M2ML abstracts the tasks into three key elements that can be easily called and accessed by users. M2ML is capable of utilizing a variety of computational resources, that are distributed across scientific facilities, to accelerate machine learning tasks. For example, it can automatically transfer data from an experimental facility (such as APS-U) to a high performance computing (HPC) facility (such as the Argonne Leadership Computing Facility (ALCF)), train machine learning models at the HPC facility, and deploy the trained models on edge computing devices back at the experimental facility for inferencing. M2ML provides a unified interface for (on-the-fly) model (re)training, storage, evaluation, fine-tuning, and inferencing using heterogeneous resources that can be geographically distributed. M2ML uses Globus services such as Globus Transfer and Globus Compute (formerly FuncX). We evaluate M2ML using a high energy diffraction microscopy (HEDM) workflow that employs BraggNN to predict the diffraction peak locations. Results show that, although the BraggNN model is small, M2ML can significantly accelerate the workflow through selective assignment of tasks to different computing resources.

Deep learning methods are transforming research, enabling new techniques, and ultimately leading to new discoveries. As the demand for more capable AI models continues to grow, we are now entering an era of Trillion Parameter Models (TPM), or models with more than a trillion parameters---such as Huawei's PanGu-Σ. We describe a vision for the ecosystem of TPM users and providers that caters to the specific needs of the scientific community. We then outline the significant technical challenges and open problems in system design for serving TPMs to enable scientific research and discovery. Specifically, we describe the requirements of a comprehensive software stack and interfaces to support the diverse and flexible requirements of researchers.

Pooling and sharing data increases and distributes its value. But since data cannot be revoked once shared, scenarios that require controlled release of data for regulatory, privacy, and legal reasons default to not sharing. Because selectively controlling what data to release is difficult, the few data-sharing consortia that exist are often built around data-sharing agreements resulting from long and tedious one-off negotiations. We introduce Data Station, a data escrow designed to enable the formation of data-sharing consortia. Data owners share data with the escrow knowing it will not be released without their consent. Data users delegate their computation to the escrow. The data escrow relies on delegated computation to execute queries without releasing the data first. Data Station leverages hardware enclaves to generate trust among participants, and exploits the centralization of data and computation to generate an audit log. We evaluate Data Station on machine learning and data-sharing applications while running on an untrusted intermediary. In addition to important qualitative advantages, we show that Data Station: i) outperforms federated learning baselines in accuracy and runtime for the machine learning application; ii) is orders of magnitude faster than alternative secure data-sharing frameworks; and iii) introduces small overhead on the critical path.

Exponentially increasing data volumes, coupled with new modes of analysis have created significant new opportunities for data scientists. However, the stochastic nature of many data science techniques results in tradeoffs between costs and accuracy. For example, machine learning algorithms can be trained iteratively and indefinitely with diminishing returns in terms of accuracy. In this paper we explore the cost-accuracy tradeoff through three representative examples: we vary the number of models in an ensemble, the number of epochs used to train a machine learning model, and the amount of data used to train a machine learning model. We highlight the feasibility and benefits of being able to measure, quantify, and predict cost accuracy tradeoffs by demonstrating the presence and usability of these tradeoffs in two different case studies.

Materials discovery is crucial for making scientific advances in many domains. Collections of data from experiments and first-principle computations have spurred interest in applying machine learning methods to create predictive models capable of mapping from composition and crystal structures to materials properties. Generally, these are regression problems with the input being a 1D vector composed of numerical attributes representing the material composition and/or crystal structure. While neural networks consisting of fully connected layers have been applied to such problems, their performance often suffers from the vanishing gradient problem when network depth is increased. Hence, predictive modeling for such tasks has been mainly limited to traditional machine learning techniques such as Random Forest. In this paper, we study and propose design principles for building deep regression networks composed of fully connected layers with numerical vectors as input. We introduce a novel deep regression network with individual residual learning, IRNet, that places shortcut connections after each layer so that each layer learns the residual mapping between its output and input. We use the problem of learning properties of inorganic materials from numerical attributes derived from material composition and/or crystal structure to compare IRNet's performance against that of other machine learning techniques. Using multiple datasets from the Open Quantum Materials Database (OQMD) and Materials Project for training and evaluation, we show that IRNet provides significantly better prediction performance than the state-of-the-art machine learning approaches currently used by domain scientists. We also show that IRNet's use of individual residual learning leads to better convergence during the training phase than when shortcut connections are between multi-layer stacks while maintaining the same number of parameters.

Open Science is key to future scientific research and promotes a deep transformation in the whole scientific research process encouraging the adoption of transparent and collaborative scientific approaches aimed at knowledge sharing. Open Science is increasingly gaining attention in the current and future research agenda worldwide. To effectively address Open Science goals, besides Open Access to results and data, it is also paramount to provide tools or environments to support the whole research process, in particular the design, execution and sharing of transparent and reproducible experiments, including data provenance (or lineage) tracking. This work introduces the Climate Analytics-Hub, a new component on top of the Earth System Grid Federation (ESGF), which joins big data approaches and parallel computing paradigms to provide an Open Science environment for reproducible multi-model climate change data analytics experiments at scale. An operational implementation has been set up at the SuperComputing Centre of the Euro- Mediterranean Center on Climate Change, with the main goal of becoming a reference Open Science hub in the climate community regarding the multi-model analysis based on the Coupled Model Intercomparison Project (CMIP).

Scientists' capacity to make use of existing data is predicated on their ability to find and understand those data. While significant progress has been made with respect to data publication, and indeed one can point to a number of well organized and highly utilized data repositories, there remain many such repositories in which archived data are poorly described and thus impossible to use. We present Skluma---an automated system designed to process vast amounts of data and extract deeply embedded metadata, latent topics, relationships between data, and contextual metadata derived from related documents. We show that Skluma can be used to organize and index a large climate data collection that totals more than 500GB of data in over a half-million files.

Big data workflows often require the assembly and exchange of complex, multi-element datasets. For example, in biomedical applications, the input to an analytic pipeline can be a dataset consisting thousands of images and genome sequences assembled from diverse repositories, requiring a description of the contents of the dataset in a concise and unambiguous form. Typical approaches to creating datasets for big data workflows assume that all data reside in a single location, requiring costly data marshaling and permitting errors of omission and commission because dataset members are not explicitly specified. We address these issues by proposing simple methods and tools for assembling, sharing, and analyzing large and complex datasets that scientists can easily integrate into their daily workflows. These tools combine a simple and robust method for describing data collections (BDBags), data descriptions (Research Objects), and simple persistent identifiers (Minids) to create a powerful ecosystem of tools and services for big data analysis and sharing. We present these tools and use biomedical case studies to illustrate their use for the rapid assembly, sharing, and analysis of large datasets.

We present a light-weight database virtualization (LDV) system that allows users to share and re-execute applications that operate on a relational database (DB). Previous methods for sharing DB applications, such as companion websites and virtual machine images (VMIs), support neither easy and efficient re-execution nor the sharing of only a relevant DB subset. LDV addresses these issues by monitoring application execution, including DB operations, and using the resulting execution trace to create a lightweight re-executable package. A LDV package includes, in addition to the application, either the DB management system (DBMS) and relevant data or, if the DBMS and/or data cannot be shared, just the application-DBMS communications for replay during re-execution. We introduce a linked DB-operating system provenance model and show how to infer data dependencies based on temporal information about the DB operations performed by the application's process(es). We use this model to determine the DB subset that needs to be included in a package in order to enable re-execution. We compare LDV with other sharing methods in terms of package size, monitoring overhead, and re-execution overhead. We show that LDV packages are often more than an order of magnitude smaller than a VMI for the same application, and have negligible re-execution overhead.

Sharing and repeating scientific applications is crucial for verifying claims, reproducing experimental results (e.g., to repeat a computational experiment described in a publication), and promoting reuse of complex applications. The predominant methods of sharing and making applications repeatable are building a companion web site and/or provisioning a virtual machine image (VMI). Recently, application virtualization (AV), has emerged as a light-weight alternative for sharing and efficient repeatability. AV approaches such as Linux Containers create a chroot-like environment [4], while approaches such as CDE [1] trace system calls during application execution to copy all binaries, data, and software dependencies into a self-contained package.

The Ophidia project is a research effort addressing big data analytics requirements, issues, and challenges for eScience. We present here the Ophidia analytics framework, which is responsible for atomically processing, transforming and manipulating array-based data. This framework provides a common way to run on large clusters analytics tasks applied to big datasets. The paper highlights the design principles, algorithm, and most relevant implementation aspects of the Ophidia analytics framework. Some experimental results, related to a couple of data analytics operators in a real cluster environment, are also presented.

We introduce and study a new class of queries that we refer to as ACMA (arithmetic constraints on multiple attributes) queries. Such combinatorial queries require the simultaneous satisfaction of arithmetic constraints on three or more attributes from different relations, and thus often involve expensive multi-join operations. Building on techniques from constraint programming, we develop preprocessing methods, algorithms, and a new constrained join operator that allow ACMA queries to be evaluated efficiently within a conventional relational database engine. We present the results of a careful performance evaluation of both our new approach and the conventional nested-loop join algorithm. Measurements of tuples read, intermediate tuples generated, and execution time shows that our approach achieves superior performance for ACMA joins.

It is increasingly common to encounter communities engaged in the collaborative analysis and transformation of large quantities of data over extended periods of time. I argue that these communities require a scalable system for managing, tracing, exploring and communicating the derivation and analysis of diverse data objects. Such a system could bring significant productivity increases facilitating discovery, understanding, assessment, and sharing of both data and transformation resources for computation, storage, and collaboration. I define a model and architecture for a virtual data grid capable of addressing these requirements. I define a broadly applicable model of a "typed dataset" as the unit of derivation tracking, and simple constructs for describing how datasets are derived from transformations and from other datasets. I also define mechanisms for integrating with, and adapting to, existing data management systems and transformation and analysis tools, as well as grid mechanisms for distributed resource management and computation planning. Finally, I report on successful application results obtained with a prototype implementation called Chimera, involving challenging analysis of high-energy physics and astronomy data.

A lot of scientific data is not obtained from measurements but rather derived from other data by the application of computational procedures. We hypothesize that explicit representation of these procedures can enable documentation of data provenance, discovery of available methods, and on-demand data generation (so-called "virtual data"). To explore this idea, we have developed the Chimera virtual data system, which combines a virtual data catalog for representing data derivation procedures and derived data, with a virtual data language interpreter that translates user requests into data definition and query operations on the database. We couple the Chimera system with distributed "data grid" services to enable on-demand execution of computation schedules constructed from database queries. We have applied this system to two challenge problems, the reconstruction of simulated collision event data from a high-energy physics experiment, and searching digital sky survey data for galactic clusters, with promising results.