Skip to Main Content U.S. Department of Energy
Computing Research

Machine Learning

Machine Learning is used to address critical national and global issues by applying scientific and mathematical techniques in artificial intelligence to multiple data sources and communicating these findings to the community. Researchers also develop novel, interpretable, and scalable machine learning algorithms and applications, designed to run on systems as large as current DOE leadership class facilities all the way to commodity hardware.

Key Capabilities

  • Extreme-scale Machine Learning, including Fault-tolerant and Communication-reducing Algorithms
  • Deep Learning, Ranging Supervised Learning to Deep Reinforcement Learning to Generative Adversarial Networks
  • Recommender Systems, from Semantic Knowledge Graphs to Representation Learning
  • Developing and Applying Machine Learning Algorithms (e.g., Power Systems, High Energy Physics, Biological Systems, Social Media Analytics, and Cyber Security)

Significant Projects

Deep Learning Website

Deep Learning

Contemporary deep learning has enabled the next generation of artificial intelligence applications, opening the door for major breakthroughs. PNNL applies deep learning across its mission sciences in energy, biology, the environment, and national security to better leverage data and computational resources to accelerate innovation and enhance scientific discovery.

MaTEx Website

Extreme-scale TensorFlow and Keras with Machine Learning Toolkit for Extreme Scale (MaTEx)

This project continues to design scalable and fault-tolerant machine learning and deep learning algorithms for leadership class facilities.

Sharkzor: Interactive Visual Sort and Summary of Imagery

Sharkzor combines large, pre-trained convolutional neural nets with much smaller interest-specific “neural modules.” This allows subtle, dynamic interaction with social media images not possible before.

Power Systems

This project investigates an ensemble-based technique, Bayesian Model Averaging (BMA), to improve Net Interchange Schedule (NIS) forecast performance. This work illustrates a possible new mechanism for improving NIS forecasting accuracy, as well as other power grid system variables.

Biosurveillance for Early Warning of Disease Outbreaks

In this set of projects, deep learning and machine learning techniques are being applied to derive insight about disease incidence from unstructured data, including news and social media.

Learning Control for Building Systems

This project is investigating, developing, and testing automated online learning control for building systems using deep learning algorithms.

GA Wiki

Global Arrays

This project is expanding Global Arrays (GA) by designing extreme-scale Exascale Computing Project applications, such as NWChemEx, GAMESS, and enhancing GA for designing next-generation machine learning and deep learning algorithms.

Exploring Fault Patterns and Approximation Impact on High-Performance Scientific Applications Through Machine Learning

This project is exploring the sensitivity of high-performance computing applications to faults and approximation, understanding performance and accuracy trade-offs, and building models to determine if an application is suitable for approximation.

Computing Research

Research Areas