Sorry, you need to enable JavaScript to visit this website.


Design Intelligence Fostering Formidable Energy Reduction and Enabling Numerous Totally Impactful Advanced Technology Enhancements

In the 250 years since the dawn of the Industrial Revolution, the pace of technology-driven economic growth has dwarfed that achieved in prior centuries. The emerging artificial intelligence revolution has similar transformational potential, which we seek to leverage to help resolve the energy and environmental challenges that are tied to the modern industrial age. Artificial intelligence (A.I.) makes it possible for machines to learn from experience, adjust to new inputs and perform like humans. Machine learning is a core part of A.I., and it is the study of computer algorithms that improve automatically through experience. Incorporating machine learning into the energy technology and/or product design processes is anticipated to facilitate a rapid transition to lower-carbon-footprint energy sources and systems.

Carnegie Mellon University

Predicting Catalyst Surface Stability under Reaction Conditions Using Deep Reinforcement Learning and Machine Learning Potentials

Carnegie Mellon University will use deep reinforcement learning and atomistic machine learning potentials to predict catalyst surface stability under reaction conditions. Current methods for determining the metastability of bifunctional and complex surfaces undergoing reaction are difficult and expensive. Carnegie Mellon's technology will enable stability analysis in both traditional catalysts and new classes of materials, including those used in tribology (friction), corrosion-resistant alloys, additive manufacturing, and battery materials.

General Electric

Pro-ML IDeAS: Probabilistic Machine Learning for Inverse Design of Aerodynamic Systems

GE Global Research will develop a probabilistic inverse design machine learning (ML) framework, Pro-ML IDeAS, to take performance and requirements as input and provide engineering designs as output. Pro-ML IDeAS will calculate the design explicitly without iteration and overcome the challenges of ill-posed inverse problems. Pro-ML IDeAS will use GE's Bayesian hybrid modeling with multi-fidelity intelligent design and analysis of computer experiments and a novel probabilistic invertible neural network (INN). The proposed framework can be applied to general complex design problems. The designs of interest are turbomachinery components, applicable to not only industrial gas turbines, but also aviation turbine engines, aero derivative engines, wind turbines, and hydro turbines.

IBM T. J. Watson Research Center

Model-based Reinforcement Learning with Active Learning for Efficient Electrical Power Converter Design

Iowa State University

Context-Aware Learning for Inverse Design in Photovoltaics

Iowa State University will develop novel machine learning tools to accelerate the inverse design of new microstructures in photovoltaics. The team will create a new deep generative model called bi-directional inverse design networks to combat challenges in real-world inverse design problems. The proposed inverse design tools, if successful, will produce novel, manufacturable material microstructures with improved electromagnetic properties relative to existing technology for better, more efficient solar energy.

Lawrence Berkeley National Laboratory

Deep Learning and Natural Language Processing for Accelerated Inverse Design of Optical Materials

Over the past 50 years, progress in optical metamaterial device design has led to the manipulation of light over a wide range of wavelengths spanning the ultraviolet to the far infrared, resulting in technological advancements such as selective radiative absorbers for solar energy and daytime passive cooling using deep space. Further advances in optical metamaterial devices could enable increased energy efficiency, reduced national primary energy consumption, inexpensive long duration energy storage, and next generation solid-state heat engines. Lawrence Berkley National Laboratory (LBNL) will develop an optical metamaterial design tool to increase energy efficiency and reduce national primary energy consumption. Besides creating high-quality datasets, LBNL will train physics-informed generative adversarial networks that automatically suggest candidate structures to produce desired optical properties within the constraints of cost of materials and manufacturing. Currently, finding an optimal design can take years and is based mostly on intuition and iteration. The team's machine learning tool will be 10,000 to 100,000 times faster than existing technology.

Massachusetts Institute of Technology

Machine Learning Assisted Models for Understanding and Optimizing Boiling Heat Transfer on Scalable Random Surfaces

The Massachusetts Institute of Technology (MIT) will develop a machine learning (ML) approach to optimize surfaces for boiling heat transfer and improve energy efficiency for applications ranging from nuclear power plants to industrial process steam generation. Predicting and enhancing boiling heat transfer presently relies on empirical correlations and experimental observations. MIT's technology will use supervised ML models to identify important features and designs that contribute to heat transfer enhancement autonomously. If successful, MIT's designs will lead to more readily adopted scalable surfaces in energy applications, enhancing performance and shortening deployment timetables.

National Renewable Energy Laboratory

End-to-End Optimization for Battery Materials and Molecules by Combining Graph Neural Networks and Reinforcement Learning

The National Renewable Energy Laboratory (NREL) will develop a machine learning-enhanced approach to the design of new battery materials. Currently, such materials are designed in part via numerous expensive high-fidelity computational simulations that predict the performance of a given composition. However, at present, humans must sift through the vast amounts of data generated and manually identify new compositions. To accelerate this process, NREL plans to develop a machine learning enhanced prediction tool that uses existing simulation data to predict the performance of new material compositions at high fidelity but lower cost. NREL plans to combine this tool with reinforcement learning techniques to automate the identification of new candidate compositions. It is expected that these design tools will enable the identification of new battery materials faster and thereby accelerate the rate at which battery performance is improving.

National Renewable Energy Laboratory

INTEGRATE - Inverse Network Transformations for Efficient Generation of Robust Airfoil and Turbine Enhancements

The National Renewable Energy Laboratory (NREL) will develop a novel wind turbine design capability that enables designers to explore advanced technology concepts at a lower cost. This capability will harness the power of a deep neural network (DNN)-based inverse design methodology. To overcome challenges with the use of traditional DNNs in this application, NREL will develop innovative techniques to sparsify the neural network using active subspaces that will ensure that the model is invertible and can quickly zoom in on relevant designs at minimal cost. The models will be trained using data from computational fluid dynamics simulations, running on NREL's supercomputers, which in turn use machine learning assisted turbulence models to predict flow separation and stall observed in wind turbine flows.

Pacific Northwest National Laboratory

Machine Learning for Natural Gas to Electric Power System Design

Pacific Northwest National Laboratory (PNNL) will apply multiple machine learning tools to develop next-generation natural gas to electric power conversion system designs. The project leverages a physics-informed machine learning tool for automated reduced order model (ROM) construction. This will significantly reduce prediction errors compared to traditional approaches. Machine learning will also leverage a superstructure-based mathematical optimization tools combined with reinforcement learning and graph network methods to explore and optimize component connections in fuel to electric power conversion systems.

Princeton University

MLSPICE: Machine Learning based SPICE Modeling Platform for Power Magnetics

The Princeton University team will use machine learning-enabled methods to transform the modeling and design methods of power magnetics and catalyze disruptive improvements to power electronics design tools. They will develop a highly automated, open-source, machine learning-based magnetics design platform to greatly accelerate the design process, cut the error rate in half, and provide new insights to magnetic material and geometry design. Princeton's Simulation Program with its Integrated Circuit Emphasis-based, or SPICE-based modeling platform, will utilize a highly automated data acquisition testbed capable of measuring a large number of magnetic cores with a wide range of electrical circuit excitations, a machine-learning trained modeling method for modeling the core loss and saturation effects of magnetic materials, and a computer-aided-design tool which can synthesize the SPICE netlist for planar magnetics.

Stanford University

Energy Efficient Integrated Photonic Systems Based on Inverse Design

Stanford University will develop a machine-learning enhanced framework for the design of optical communications components that will enable them to operate at their physical performance limits. Information processing and communications systems use a significant fraction of total global energy. Data centers alone consume more than 70 billion kilowatt-hours per year. Much of this energy usage is intrinsic to electronic wiring. However, optical-based technologies offer a promising option to reduce energy consumption. Stanford's design platform is intended to enable optical technologies to serve in the next generation of information processing hardware with ultra-low energy footprints. The proposed framework will use generative neural networks for global optimization of nanophotonic components, machine learning to accelerate the solving of electromagnetic field calculations, and advanced optimization concepts to calculate the upper limits in photonic device performance.

University of Missouri

Deep Learning Prediction of Protein Complex Structures

The University of Missouri will develop deep learning methods to predict inter-protein amino acid interactions and build three-dimensional structures of protein complexes, which are useful for designing and engineering protein molecules important for renewable bioenergy production. Proteins in cells interact and form complexes to carry out various biological functions such as catalyzing biochemical reactions. The team will use the deep learning methods it develops to construct green algae protein complexes that play important roles in biomass and biodiesel production. The technology and predicted structures of protein complexes will become valuable tools and resources for advancing U.S. bioenergy production and research.