BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Europe/Stockholm
X-LIC-LOCATION:Europe/Stockholm
BEGIN:DAYLIGHT
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
TZNAME:CET
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20210518T152203Z
LOCATION:Substanz 1\, 2
DTSTART;TZID=Europe/Stockholm:20190619T083000
DTEND;TZID=Europe/Stockholm:20190619T090000
UID:isc_hpc_ISC High Performance 2019_sess226_pap172@linklings.com
SUMMARY:MaLTESE: Large-Scale Simulation-Driven Machine Learning for Transi
 ent Driving Cycles
DESCRIPTION:Research Paper\n\nMaLTESE: Large-Scale Simulation-Driven Machi
 ne Learning for Transient Driving Cycles\n\nAithal, Balaprakash\n\nOptimal
  engine operation during a transient driving cycle is the key\nto achievin
 g greater fuel economy and reduced emissions.  A  large-scale simulation-d
 riven\nmachine-learning  approach  is  presented  in  this  work to explor
 e the parameter space for optimal engine performance and emissions.  A  pa
 rallel,  fast,  robust,\nphysics-based reduced-order engine simulator is u
 sed to obtain performance and\nemission characteristics of engines over a 
 wide range of control parameters under\nvarious transient driving conditio
 ns (drive cycles). We scale the simulation up to\n3,096 nodes of the Theta
  supercomputer at the Argonne Leadership Computing\nFacility to generate d
 ata required to train a machine learning model. The trained\nmodel is then
  used to predict various engine parameters of interest and compared\nwith 
 results predicted by the engine. Our results show that a deep neural-netwo
 rk-\nbased surrogate model achieves a high accuracy: Pearson product-momen
 t cor-\nrelation coefficient values larger than 0.99 and mean absolute per
 centage error\nwithin 1.07% for various engine parameters such as exhaust 
 temperature, exhaust\npressure, nitric oxide, and engine torque. Once trai
 ned, the deep-neural-network-\nbased surrogate-model is fast for inference
 : it requires about 16 microseconds for\npredicting the engine performance
  and emissions for a single design configuration\nas compared with about 0
 .5 second per configuration with the engine simulator.\nMoreover, we demon
 strate that transfer learning and retraining can be leveraged\nto incremen
 tally retrain the surrogate model to cope with new configurations that\nfa
 lls outside the training data space.\n\nPasses: Conference Pass\n\nTag: Co
 mmunication Optimization, MPI, Networks, Programming Models & Languages, S
 ystem Software & Runtime Systems
URL:https://2019.isc-program.com/presentation/?id=pap172&sess=sess226
END:VEVENT
END:VCALENDAR

