The ICEnet consortium

The transportation sector accounts for roughly 14% of global greenhouse gas (GHG) emissions and about 29% of GHG emissions in the U.S., per an EPA 2017 study. Given the looming threat of the climate crisis, solely focusing on improving electric vehicle technology and their adoption will not be sufficient in itself, since Internal Combustion Engines still account for over 99% of the global on-road vehicle population. It is therefore imperative that enterprising research in improving engine combustion is given importance for the foreseeable future.

ICEnet, a data-driven consortium based out of UMass Amherst, is at the forefront of this challenge. This collaborative venture utilizes the disruptive technology of artificial intelligence (AI), applying it to the engine research and development paradigm. At ICEnet we develop predictive machine learning algorithms to improve engine design cycles using computer models, allowing manufacturers to optimize their designs and create fuel-efficient engines.

Some of our current research areas include improving turbulence model fidelity, developing better sub-models for spray/gas interactions and using artificial neural networks for making combustion calculations faster (combustion calculations take up the most wall time for an engine R&D calculation)

Apart from UMass Amherst, the ICEnet consortium has among its partners leading members of the engine R&D software providers as well as leaders from the AI computing world. The list of current members include:

• AVL
• Convergent Science, Inc.
• Cummins Inc.
• German Research Center for Artifical Intelligence (DFKI)
• MathWorks Inc.
• NVIDIA Corp.
• SIEMENS/CD-adapco
• University of Massachusetts Amherst

The ICEnet project also collaborates with U.S. Department of Energy laboratory, Los Alamos on a limited basis for the turbulence phase of the project

ICEnet’s machine learning technology developed in conjunction with NVIDIA and MathWorks, is actively tested and implemented within industry-leading engine design software makers such as SIEMENS, AVL, Convergent Science (together these three account for 100% market share for engine R&D simulation platforms) and used on Cummins design cycles, provides a pathway to quickly disseminate our knowledge to end-users in the market. So far ICEnet has raised over $350,000 from partners and our early results will be presented at peer-reviewed machine learning conferences including NeurIPS 2019.

Additionally, ICEnet is diversifying its impact portfolio by reimagining the external aerodynamic design of vehicles, another determinant of fuel efficiency. ICEnet is collaborating with Carnegie Mellon University to replicate our success providing opportunities for even electric vehicle manufacturers to utilize our research.

• Turbulence Forecasting via Neural ODE, to appear, NeurIPS ML4PS 2019, Vancouver, BC, Canada
• A data-driven approach to modeling turbulent flows in an engine environment, to appear, American Physical Society DFD 2019, Seattle, WA
• A data-driven approach to modeling turbulent decay at non-asymptotic Reynolds numbers, to appear, American Physical Society DFD 2019, Seattle, WA
• Climate sub-closures using Neural ODE, talk submitted to GTC 2020
• Continuous time networks for two-equation turbulence models, poster submitted to GTC 2020
• LES Turbulence modeling with Machine Learnt Closure Model for ICEs, paper submitted to SAE WCX 2020

• Dec 2019: First ICEnet Face-to-Face meeting to be organized at UMass Amherst between Dec 2-4, 2019
• Oct 2019: ICEnet's paper to NeurIPS ML4PS 2019 has been accepted
• Sept 2019: ICEnet's paper to APS DFD 2019 has been accepted
• Aug 2019: ICEnet conducts its first quarterly update
• June 2019: ICEnet conducts its first mid-quarterly update
• May 2019: The two-year ICEnet project officially kicks off

• Partners of the consortium can access trained networks, codes, internal presentations and video recordings using the ICEnet Hub

• Spatio-temporal modeling of cycle-to-cycle variability in ICEs