Four potential high-impact projects intended to solve critical energy-efficiency challenges have been awarded a total of $300,000 in seed funding from UC Santa Barbara’s Institute for Energy Efficiency (IEE), the College of Engineering’s interdisciplinary research center dedicated to cutting-edge science and technologies that support an energy-efficient and sustainable future. Each of the projects aligns with at least one of the institute’s key interdisciplinary thrusts: smart societal infrastructure, computing and communications, and the food-energy-water nexus.
Two of the projects were the inaugural recipients of $100,000 grants from The Investment Group of Santa Barbara (IGSB), though a program launched this year to support high-impact research of energy-efficient software. The funding provides a significant advance or reduction in risk of work that is likely to lead to commercialization and a positive impact on society. Two projects received $50,000 apiece through the IEE Research Seed Grant Program. Seed funding is intended to help researchers produce preliminary results that can be used to apply for major external funding to expand their projects.
“Seed grants are an essential step in the creation and delivery of impactful solutions to improve efficiency,” said IEE director John Bowers, a distinguished professor of electrical and computer engineering and materials. “Each project attacks a grand challenge with an innovative path to discovery. The grants also foster new research collaborations and continue IEE’s legacy of leveraging the university’s highly acclaimed faculty to drive leading-edge discoveries and scientific advancements.”
The four projects are aimed at developing novel technologies to improve the efficiency of mobile devices, batteries, cryptocurrency, and management of the power grid. They involve seven faculty members and one postdoctoral scholar from the Departments of Chemical Engineering, Computer Science, Environmental Studies, and Materials. Summaries of the projects are included below.
Efficient AI Software on Mobile Devices
Modern artificial intelligence (AI) technologies are driven by powerful deep neural networks, which are composed of a massive number of simple computing units called artificial neurons. These neurons behave similarly to those in the human brain, as each one performs simple calculations such as multiplying and adding numbers. Connected in a neural network, these artificial neurons can do powerful “intelligent” tasks, such as when an application translates a sentence from English to Spanish or a car’s camera recognizes a pedestrian.
AI models are, however, gigantic, and as they become more accurate, they become even larger. A single model can contain more than one-hundred billion parameters, requiring large computer clusters to do the computations, which consume a lot electric power and produce a large carbon footprint. For example, the carbon emission of training the GPT3 AI model is equivalent to that of driving a common passenger car for one-hundred-fifty years. AI’s large energy requirements make it challenging to run AI-enabled software on mobile devices, which have less memory and less powerful chips than computers. Mobile devices can also produce excessive heat while running such software, which limits battery life.
Two computer scientists, Lei Li and William Wang, are collaborating on a project, “Compact and Energy-efficient AI Software on Mobile Devices,” to improve the performance, and reduce the size, and energy consumption of an AI model on mobile devices. Recipients of a $100,000 IGSB Software Impact Grant, they expressed gratitude to IEE and IGSB for fostering a collaborative environment and empowering their work.
“Our goal is to develop compact, energy-efficient AI software tailored for mobile devices,” said Li, an assistant professor of computer science who previously worked as the director of ByteDance AI Lab, the company behind TikTok and other prominent apps. “Right now, the actual computation of an AI model happens on the remote cloud sever. We would like to develop software that can automatically fit AI models and bring such computation to mobile devices.”
The two say that they have built a strong foundation, together and individually, to tackle this grand challenge. In addition to having collaborated previously on multiple successful projects, both researchers have extensive experience in developing machine learning for natural language understanding, text generation, machine translation, reasoning, and visual understanding. Li is the leader for LightSeq, an open-source library for high-performance computing for Transformer models on cloud-based graphics processing units (GPUs) that has already attracted 2,300 stars, or “likes” and “follows,” on GitHub. Wang develops HULK, an energy-efficiency benchmark platform for responsible natural language processing (NLP). He has also published more than eighty papers in top journals and at leading NLP/AI/Machine Learning conferences.
Li and Wang will investigate three key areas: model compactness, energy efficiency, and memory usage. They will use and collect statistics from applications on devices, and also develop adaptive algorithms to compress models and hardware-acceleration techniques to facilitate on-device AI computing. They also plan to create an open-source software library for accelerated neural computing on mobile devices and a suite of neural models developed for on-device AI applications. Their endeavor will be guided by the key insight that an energy-efficient AI system should adaptively manage the computing resources to accommodate the model architecture, hardware specifications, such as memory, central processing units (CPUs), and GPUs, and application tasks.
“We want to enable big AI models to run on small and less powerful devices while retaining high accuracy. This would greatly reduce the energy consumption for AI computation,” said Wang, the founding director of UCSB’s Center for Responsible Machine Learning and a recipient of an Early CAREER Award from the National Science Foundation and a DARPA Young Faculty Award. “Our long-term vision is to build high-quality, economic, and green AI systems, so that every researcher and developer will be able to use our software to develop their mobile phone products.”
Li and Wang hope to drive the next generation of AI technology in terms of intelligence-per-watt consumption, a measure of the energy efficiency of a particular computer architecture or hardware. In the end, they say, their work could benefit users around the world with better access to greener AI technologies.
Improving the Energy Efficiency of Cryptocurrencies
Cryptocurrency, or crypto, is a digital-payment system that does not rely on banks to verify transactions. Cryptocurrency received its name because it uses encryption, or advanced coding, to verify transactions and provide security and safety. Units of crypto are created through a process called mining, which involves using high-performance computers and an agreed upon “mining” algorithm to solve complicated mathematical problems that generate coins, a token of encrypted data that denotes a unit of value. Holders of crypto do not own anything tangible, but their ownership is recorded on a tamper-proof and unforgeable public ledger, called a blockchain, which is available to all market participants. As a result, it is possible for anyone to verify that the seller or buyer indeed owns the coin that is being purchased or presented for payment. The current market capitalization of all cryptocurrencies in circulation is roughly $920 billion, similar to the gross domestic product of Turkey, the seventeenth largest economy in the world.
A major problem with cryptocurrencies is that they are extremely inefficient and significant contributors to global air pollution. As the market has grown over time, the problems that crypto “miners” had to solve in order to mint, or earn, new coins became increasingly complex, requiring more computing power, which used more energy. Because mining is a highly competitive business, firms operate thousands of high-tech computers around the clock. Bitcoin, the world’s largest cryptocurrency, currently consumes an estimated 156 terawatt-hours of electricity annually, more than the entire country of Argentina, which is home to 45 million people. Producing that much energy emits 67 megatons of carbon dioxide annually, comparable to the emissions of New Zealand.
“We are thrilled and grateful to receive this software impact grant to study the next generation of energy-efficient cryptocurrencies and smart contracts,” said Krintz, referencing smart contracts, which are pieces of code that perform computations based on the inputs and outputs of transactions. “This is a very important area of energy efficiency, and our group is excited and eager to contribute new advances that enhance the sustainability of the next generation of computing systems.”
As part of their project, titled “Coloring Cryptcoins Green: Making Cryptocurrencies More Energy Efficient,” the two will develop a methodology by which market participants in a cryptocurrency market are incentivized to use renewable energy for their electricity expenditure while mining digital coins.
“We plan to “color” each coin minted in proportion to the renewable energy its mining used,” explained Wolski. “Transactions that use coins minted with renewable energy will be “colored green” in a tamperproof way — allowing people to choose coins and participate in transactions based on their greenness.”
Krintz and Wolski plan to access renewable energy data from Independent System Operations (ISOs) to tag each coin with an estimate of the renewable energy that was available from the grid at the time the coin was minted in the region where the verifier, which is a tool used to audit and verify each transaction, was located. In their approach, the instantaneous amount of renewable energy available in the region where the verifier is located would serve as a form of “carbon license,” incentivizing verifiers to participate when renewable energy is present. They aim to design a transaction that correctly records the relationship between a verifier and an ISO so that the available energy is correctly reflected when the coin is minted, which would make it possible to determine the fraction of total renewable energy that each minting consumed.
“Our work will reduce the energy consumption of cryptocurrency by incentivizing the use of green cryptocoins,” said Krintz, adding that the Environmental, Social, and Governance (ESG)-relevance of their work could make commercialization possible.
Their research could also spur the creation of “green contracts,” in which a smart contract can be executed only by using cryptocoins that are sufficiently green, further dropping the energy consumption of computing.
A Model of Collaboration
Lithium (LI)-ion batteries have a wide range of applications from electric vehicles and smartwatches to cellphones and laptops. Some of their advantages over other rechargeable battery technologies include their higher energy density and voltage, as well as their lack of a memory effect, a detrimental process that causes a battery to “remember” a lower capacity after repeated charge cycles.
However, this technology is approaching its practical performance limits, spurring efforts to find new electrochemical storage solutions that are more stable and economically viable, age better, and possess higher energy density, always the holy grail of battery performance. Using a metallic Li in place of conventional graphite-based anodes is a promising solution for researchers, because it would increase the overall energy density of Li-ion batteries. Unfortunately, there are many challenges in using metallic Li, arising from complex electrochemical phenomena that result in reduced storage capacity losses, rapid degradation, and safety issues. Two modeling experts from UCSB, materials professor Anton Van der Ven, and chemical engineering professor M. Scott Shell, have received an IEE seed grant to address the multi-disciplinary challenges posed by incorporating metallic Li into energy-storage devices. They hope to unravel the fundamental mechanisms that drive the complex phenomena that occur within batteries.
“We currently have a limited ability to understand the complex electrochemical phenomena and the transformations that they incur within a battery by purely experimental methods,” said Van der Ven. “We believe that recent advances in machine learning and computational modeling have created exciting opportunities in materials science that will enable significant breakthroughs in energy storage and conversion.”
In their project, he and Shell will develop a software infrastructure to enable atomistic simulations of batteries and other materials based on higher-resolution first principles electronic structure theory that describes fundamental interactions with quantum mechanical accuracy. The software would also allow for the modeling of poorly understood meso- and macroscopic-length-scale phenomena on which electrochemical processes of batteries occur. The software will incorporate novel machine learning approaches to learn interaction models from the quantum chemical calculations, which will then enable accurate molecular simulations at larger, supra-atomic scales.
“This collaboration cross-pollinates the best of recent advances in both solid and soft material simulations,” said Shell. “What is so exciting is that this project combines state-of-the-art techniques for modeling a solid material’s detailed electronic structures, which require quantum resolution, with multiscale simulation techniques used to model modern soft materials that will be able to translate these interactions to a larger scale and determine functional behavior and properties.”
Shell and Van der Ven bring a wealth of experience and expertise to the project. Van der Ven’s group developed the CASM software suite, an open-source code base to automate first-principle statistical mechanical predictions of thermodynamic and kinetic properties of solid materials with highly defined and repeatable arrangements of molecular chains, called crystalline materials. Shell, a leader in the fields of molecular simulation, multiscale modeling, and statistical thermodynamic approaches, says that the project is an exciting bridge between two typically separate fields of research, a testament to UCSB’s unique interdisciplinary research model.
“This kind of seed funding proves incredibly enabling for what UCSB does best — linking fields and provoking new interactions at atypical intersections of expertise to solve difficult problems,” he said.
The researchers say that their software package will provide first principles insight about poorly understood non-equilibrium phenomena in batteries that are currently preventing significant increases in energy density and lifetimes. On a larger scale, the software could be applied to other poorly understood materials and establish UCSB as a leader in a movement to leverage machine learning in materials property prediction and new material discovery.
Forecasting a Greener Future
To ensure that the nation’s electricity grid operates reliably, and the lights stay on, electricity supply always need to equal demand. That supply comes from many sources, including hydro, nuclear, coal, natural gas, and, increasingly renewable energy, such as wind and solar. Unlike conventional generation, however, which can be controlled by changing the amount of fuel or water supplied to a power plant, wind and solar generation depend on when the wind blows or when the clouds clear. In order to meet power demand, electricity-grid operators have to be able to forecast how much wind or solar radiation will be available the following day and how much of that will be translated to electricity generation.
Electricity-grid operators currently use deterministic forecasts, which means they receive a single value to represent a forecast for any moment in time. Based on these forecasts, they keep some amount of conventional generation and battery storage capacity aside as reserves, in case of forecast errors, which result in significant costs. One proposed solution to save money for utility companies and, ultimately, consumers is a probabilistic forecast, which allows grid operators to change the reserve requirements based not only on the forecast, but also on the probability of being wrong. Since both electricity demand and renewable-energy generation are subject to the next day’s weather conditions, both variables factor into the decision-making process. A pair of UCSB researchers, Guillermo Terrén-Serrano and Ranjit Deshmukh, have received an IEE seed grant to develop a probabilistic method to forecast not only wind and solar generation, but also electricity demand for the next twenty-four hours.
“We are incredibly excited to receive support from the IEE for researching and contributing to our vision to solve a problem that affects us daily, which is to maintain the reliability of our electricity supply,” said Terrén-Serrano, a postdoctoral scholar in UCSB’s Clean Energy Transformation Lab (CETlab) and the Environmental Markets Lab (emLab), who earned his PhD in electrical and computer engineering from the University of New Mexico. “California recently experienced a week-long heat wave that set energy demand records. Limiting carbon emissions produced by conventional thermal energy resources, such as natural gas and coal, is critical to mitigate the effects of climate change. But non-carbon sources like wind and solar depend on weather and are variable and uncertain, a challenge that we aim to address through our research.”
Their project is aimed at decreasing the risk of power shortages associated with renewable energy by reducing weather-based uncertainty. Terrén-Serrano and Deshmukh say that a probabilistic approach is critical to determining not only what is most likely to happen, but also the probability of an extreme scenario happening tomorrow given current weather conditions.
“Our overall objective is to increase the participation of renewable energy in modern power grids,” said Deshmukh, an assistant professor in the Environmental Studies Department and an affiliated professor in the Electrical and Computer Engineering Department. “Scaling up renewable-energy generation in electricity grids will significantly mitigate carbon emissions, but we need first to address the challenge of maintaining the reliability and resiliency of our future electricity grid with high shares of renewable energy.”
The $50,000 in seed funding will allow the team to download and analyze data sufficient to define their methodology, formulate algorithms and models, perform experiments, review their results, and write academic articles to maximize the impact of their work.