Research

My research develops AI/ML-driven control and prediction frameworks for electrified powertrains, bridging physics-based modeling with modern machine learning to improve vehicle efficiency, reliability, and sustainability. Below are my primary research themes.

Aging-Aware Reinforcement Learning for Powertrain Control

I develop deep reinforcement learning frameworks (SAC, PPO, GRPO) that jointly optimize fuel economy, emissions, and component longevity in hybrid-electric vehicles. By embedding physics-based aging penalties into the RL reward structure, these controllers learn to balance short-term efficiency with long-term reliability.

Key results:

  • Fuel consumption within 1.1–4.8% of the dynamic programming (DP) optimum
  • Battery cycling degradation reduced by up to 88%
  • Generator lifetime loss reduced by up to 26.3%
  • Validated on heavy-duty Class 8 series-hybrid drive cycles in collaboration with Cummins Inc.

Key contributions:

  • Physics-Aware SAC (PA-SAC): Multi-objective energy management integrating real-time aging feedback into the RL state and reward
  • Critic-Free Policy Optimization (A-GRPO): Eliminates the critic network using trajectory-level advantage ranking with Transformer policies for long-horizon constrained control
  • Sequence-Aware Architectures: GRU and Decision Transformer-based policies for temporal reasoning in energy management
  • DP Benchmarking Framework: Dynamic Programming as a reference for RL policy evaluation and expert trajectory generation to accelerate training

Physics-Informed Transformer for Battery Aging Prediction

Accurate battery degradation prediction is critical for onboard health management and warranty planning. I developed a Physics-Informed Transformer Model (PITM) that integrates reduced-order physics-based aging trends with multivariate cycling data.

Key results:

  • Prediction error (RSEP) as low as 2.96% across 17 cells from 5 experimental datasets
  • Outperforms LSTM and standard Transformer baselines
  • Generalizes across diverse cycling conditions and battery chemistries
  • Computationally feasible for real-time deployment

Approach:

  • Combines physics-based capacity fade features (SEI growth, active material loss) with data-driven Transformer attention mechanisms
  • Trained on multi-cell datasets to learn transferable degradation patterns rather than cell-specific behavior

Physics-Based Aging Models for Powertrain Components

I build, calibrate, and validate reduced-order aging models for three critical powertrain subsystems, integrated into a forward-looking MATLAB/Simulink powertrain simulator for heavy-duty Class 8 series-hybrid electric vehicles.

Components modeled:

  • Battery: SEI layer growth and active material loss under varying C-rates, temperatures, and SOC windows
  • Electric Machine: Thermal degradation of insulation and permanent magnets under real-world load profiles
  • Aftertreatment: Catalyst deactivation and conversion efficiency loss over extended operation

Application: System-level aging analysis over real-world driving cycles, enabling aging-aware control and fleet-level lifetime management decisions.

Multi-Objective Optimization Under Uncertainty

My M.S. thesis investigated how parametric uncertainty affects non-dominated Pareto fronts in multi-objective optimization.

Key contributions:

  • Derived analytical expressions for the probability that a solution remains non-dominated under uncertainty
  • Developed a partition-based algorithm using weighted entropy to progressively reduce uncertainty in the decision space
  • Established the distribution of Frechet distances of Pareto fronts and quantified how it changes with progressive uncertainty reduction
  • Applicable to problems in energy, aging, emissions, and fleet routing optimization

Human Reliability Assessment for Nuclear Security

During my earlier work with Dr. Carol Smidts, I contributed to research on human behavior under extreme situations at nuclear facilities.

Key contributions:

  • Designed a VR-based experimental framework for collecting human performance data under physical security threats
  • Identified and measured factors influencing human behavior in extreme situations
  • Integrated experimental data into an extended THERP (Technique for Human Error Rate Prediction) framework
  • Contributed to a multi-criteria sensor selection framework for nuclear facility online monitoring systems using NSGA-based optimization