The following article is part of an ongoing series on Argonne National Laboratory’s efforts to use the predictive power of machine learning to advance discoveries in a broad range of scientific disciplines. This series will highlight the wide-ranging use of machine learning techniques across the lab, focusing on specific domains that are using this computational method to accelerate research.
Consider just a few of the major challenges in transportation — traffic congestion, pollution, inefficiencies in energy use. For decades, researchers have hammered away at these problems, making major inroads with the help of big data and high-performance computing. Now, they have a new tool in their arsenal — machine learning.
A form of artificial intelligence — a branch of computer science dealing with the simulation of intelligent behavior in computers — machine learning helps researchers discover hidden patterns within data. Using large sets of verified data, known as training data, machine-learning algorithms are taught to identify relationships between inputs and outputs. The algorithm then generates a model that can be used to make predictions on new data.
Researchers at the U.S. Department of Energy’s (DOE) Argonne National Laboratory are leveraging machine learning to transform America’s transportation and energy systems. Improved models of complex mobility and energy-related processes deepen our understanding of engine efficiencies and urban transportation needs while providing decision-makers with the data from those models that represent reality.
Transportation bottlenecks — portions of roadways that significantly affect transportation systems — are a huge cost to consumers and the economy, contributing to millions of hours and dollars wasted in idle traffic. While their impact is clear, their cause is not. Any number of diverse and complex reasons can trigger disruptions: vehicle accidents, construction, disruptive terrain, weather, limited road capacity or even badly timed traffic lights.
The rise of technologies connecting humans and vehicles — automation and “smart” infrastructure — has increased the quantity and quality of transportation data available for studying traffic states and disruptions. With more data, opportunities expand for data-driven analysis, pattern recognition and prediction for transportation systems.
While large-scale datasets offer promise, the data that inform them are often quite complex. Most data from transportation systems include significant amounts of spatially and temporally correlated information that pose a variety of challenges to traditional machine-learning approaches.
Research engineer Eric Rask, computer scientist Prasanna Balaprakash and postdoctoral researcher Tanwi Mallick are investigating ways to overcome these challenges through emerging machine-learning techniques. The team is pursuing the research through a project funded by the Vehicle Technologies Office within DOE’s Office of Energy Efficiency and Renewable Energy that uses high-performance computing resources at the Argonne Leadership Computing Facility (ALCF), a DOE Office of Science User Facility.
Rask and Balaprakash will develop and apply machine-learning algorithms to identify patterns and predict traffic from real vehicle and infrastructure data. Once implemented, these techniques will ingest data such as traffic flow, weather and road conditions, terrain, construction, vehicle accidents and more to improve estimates of future traffic flows and identify — and correct — conditions that precede a major slowdown or incident.
“To identify and fix transient problems within these networks, we first have to understand how the network behaves. High-performance computing enables us to process the vast amount of data collected about these systems, and machine learning helps us to accelerate our understanding of these problems,” said Balaprakash.
Down the road, this predictive capability, partnered with smart technologies at the vehicle and infrastructure level, could be used by consumers and transportation system operators, to support decision-making.
For example, if travelers could anticipate the formation of a bottleneck on the road, they could pre-emptively choose alternate routes to their destination. Similarly, emergency coordinators could optimize their responses to emergencies based on estimations of where bottlenecks are likely to occur. Urban planners and transportation engineers can also put in place systematic controls to mitigate bottlenecks or eliminate them altogether.
Machine-learning techniques used at the system level can also be used at the vehicle level for smarter routing, energy analysis and more. Smarter choices can help minimize fuel usage, emissions or time, but they are difficult to pinpoint due to the multitude of routing and transit options available, particularly in larger cities.
Making sense of routing options and their associated energy, time and environmental costs can help travelers and fleet operators optimize their vehicle and route selection. However, doing so requires accurate energy information and reliable predictions.
“You can get this with high-fidelity simulations, but machine learning is another option for getting an acceptable answer right away,” said Argonne Vehicle and Mobility System manager Aymeric Rousseau.
Rousseau and his team are exploring the use of machine learning to optimize routing for fleets. They start by gathering data on the various factors influencing travel, such as road and weather conditions and vehicle performance, and use these to develop reliable predictive models.
Rousseau and his team use a similar process to help policymakers predict the energy impact of emerging technologies, including engine, transmission, lightweighting and electric drive technologies. To estimate energy impacts, individual vehicles must be simulated to represent every combination of vehicle, powertrain and component technology — a process that is computationally intensive and time consuming.
“While Argonne has developed processes to individually model and simulate close to 1.5 million possible vehicle combinations, many more options are still possible,” Rousseau said. “Using machine-learning models trained from simulation results speeds up the process, allowing us to quickly answer stakeholders’ questions.”
Rather than simulating each of the millions of possible vehicle technology combinations, researchers simulate only a fraction of them, using the data generated to train and later validate machine-learning models. These models are then used to inform decision-making by helping stakeholders understand, for example, the impact a specific engine will have on a vehicle’s overall energy consumption and performance.
A more efficient engine can deliver better fuel economy, lower emissions and a greater range of performance to consumers everywhere. By using machine learning to accelerate the design of engines and combustion models, researchers at Argonne contribute to the development of better engines and shorten their path to commercialization.
Among their recent successes is the application of a method called deep learning to create a new combustion model that reduces simulation time by half.
Deep learning uses a class of algorithms called deep neural networks that mimic the brain’s simple signal processes in a hierarchical way; today, these networks, aided by high-performance computing, can be several layers deep. They enable researchers to model increasingly complex, interconnected processes like the multiple reaction pathways in fuel combustion.
“Traditionally, researchers will try to reduce the complexity of combustion reactions to save time when running simulations, but doing so can reduce the accuracy of their output,” said Argonne’s Computational Multi-Physics Section’s manager Sibendu Som.
“With our new model, aided by machine learning, we can account for the entire fuel chemistry without sacrificing accuracy,” he added. “This capability is unique, not only in its application of neural networks, but also in its ability to significantly reduce development time.”
Machine learning also played a hand in helping Argonne researchers optimize an engine designed to run on a new fuel. This work, done in collaboration with Aramco, a global petroleum company, and Convergent Science, drastically reduced model run time from months to days.
Traditional approaches for optimizing a new product design are lengthy, involving continuous experimental testing and evaluation. But by augmenting high-fidelity simulation with machine learning, Argonne was able to speed up the overall design process.
High-fidelity simulations performed at the ALCF were critical to the process. Thousands of engine design combinations simulated on the Mira supercomputer provided the data needed to train the machine-learning model to accurately predict engine performance characteristics.
From large, complex systems to individual components, machine learning renews hope that transportation problems that once seemed intractable can be solved. By continuing to grow its machine-learning competencies, Argonne is at the forefront of putting those solutions within reach.
The Office of Energy Efficiency and Renewable Energy supports early-stage research and development of energy efficiency and renewable energy technologies to strengthen U.S. economic growth, energy security, and environmental quality.
Argonne National Laboratory seeks solutions to pressing national problems in science and technology. The nation's first national laboratory, Argonne conducts leading-edge basic and applied scientific research in virtually every scientific discipline. Argonne researchers work closely with researchers from hundreds of companies, universities, and federal, state and municipal agencies to help them solve their specific problems, advance America's scientific leadership and prepare the nation for a better future. With employees from more than 60 nations, Argonne is managed by UChicago Argonne, LLC for the U.S. Department of Energy's Office of Science.
The U.S. Department of Energy's Office of Science is the single largest supporter of basic research in the physical sciences in the United States and is working to address some of the most pressing challenges of our time. For more information, visit https://energy.gov/science