Automakers looking to more quickly understand and mitigate combustion cycle variability to improve fuel economy and lower pollutant emissions are finding their magic bullet in the computational horsepower and advanced data analytics of the U.S. Department of Energy national labs. Testing that once took months can now take minutes, yielding information that enables engineers to focus more sharply on the factors causing inefficient fuel combustion.
In one such partnership, a major auto manufacturer has partnered with Argonne National Laboratory to test the company’s piston, cylinder, and other fuel system component designs. The study brings cutting-edge technology to bear on the age-old challenge of cycle-to-cycle variability in spark-ignition engines.
“Understanding the causes of cycle-to-cycle variability is critical for developing strategies to mitigate it,” said Sibendu Som, project leader for Argonne and head of the Computational Fluid Dynamics team at Argonne’s Center for Transportation Research. “It gives us options for how to control and eventually reduce it.”
Randomness over time demands that researchers simulate engine combustion instability over hundreds of consecutive cycles. Argonne uses machine learning (ML), in the form of automated model building, to understand which combinations and values of complex phenomena prompt fuel combustion cycles to veer off course, so car manufacturers can design engine components that maximize stability.
Many factors, including port design, fuel octane rating, cylinder pressure, air/fuel ratio, engine load and ignition timing, influence the stability of a cylinder’s working cycle. Researchers must simulate which of hundreds of possible scenarios lead to engine inefficiency, often manifested as engine “knock.” In addition to triggering environmental concerns, knock often heralds catastrophic damage to a car’s engine.
“It is impossible to design knock controllers using modeling systems that don’t account for randomness,” explained Muhsin Ameen, a research engineer in Argonne’s Energy Systems division. “A single-time history simulation or experiment of knock-control methods cannot provide a reliable measurement of the controller’s performance because of the random nature of arriving knock events. Performing thousands of simulations using conventional computational fluid dynamics can be extremely expensive; hence the application of ML systems.”
ML systems can learn from data, identify patterns and make decisions to create predictive models of the relationships between cyclic variability and design parameters. Center for Transportation Research scientists are creating sophisticated programs that feed simulation results back into Argonne’s powerful supercomputers to enable machine development of models to predict cyclic variability.
“The goal of this work is to exploit the unique capabilities of machine learning to uncover the complex relationships among the in-cylinder flow-fields, initial flame kernel development, and the eventual performance of the engine cycle,” said Pinaki Pal, a research engineer in the Energy Systems division. “This will allow us to develop a better understanding of random combustion phenomena leading to cyclic variability. Machine learning is advancing engineering for the cars of the future.”