Mechanical Engineering
Graduate Seminar
NOTICE OF SEMINAR PRESENTATION
CANDIDATE: Ahmad Mohammad
DEGREE SOUGHT: MASc
DATE: 7/10/2026
TIME: 11:30am
PLACE: Room 1101 CEI
TITLE: A Systematic Approach for Digital Twinning a Dual-Motor Electric Vehicle Powertrain Controller Using Neural Networks
Abstract
As electrification continues to expand in the automotive industry, model-based engineering (MBE) becomes an increasingly useful methodology for the design and benchmarking of new powertrain systems. Although physical electric vehicle (EV) subsystems can be modeled through physics-based approaches, the supervisory control logic remains proprietary. This is especially challenging in dual-motor powertrains, where the front-rear torque-allocation strategy is not directly available to the modeler. This study presents a methodology for digital twinning of a dual-motor EV powertrain when the controller strategy is unknown. A physics-based model of a 2023 Hyundai Ioniq 5 dual-motor EV was developed, including the battery, front- and rear-motors, driveline, wheels, longitudinal vehicle dynamics, and driver-demand subsystem. Simulated battery voltage, current, and state of charge (SOC) were compared against experimental drive-cycle data to validate the model. A rule-based controller was first used for baseline closed-loop validation. Subsequently, two neural-network candidates, the time-delay neural network (TDNN) and the long short-term memory (LSTM), were investigated to approximate the hidden front-rear torque allocation. Both controllers were trained and validated using the Urban Dynamometer Drive Schedule (UDDS) and Highway Fuel Economy Test (HWFET) data, then tested in closed loop on the New European Driving Cycle (NEDC) and Worldwide Harmonized Light Vehicles Test Procedure (WLTP). Performance was evaluated based on root-mean-squared errors (RMSEs) of front- and rear-motor torques, HV battery voltage, HV battery current, vehicle speed, and SOC. Both neural-network controllers successfully reproduced the torque-allocation behavior, but the TDNN provided the stronger overall closed-loop performance.