Major Research Projects
Probabilistic System Modeling and Adaptablity for Engine-out NOx [Jan 23 - Present]
Developed a predictive probabilistic model for modeling engine-out NOx using gaussian process regression.
Incorporated physical laws in the deep kernel with a causal graph derived via graph neural networks.
Quantified the epistemic and aleatory uncertainty in NOx predictions.
Assessed the performance of the model on various quantitative and qualitative metrics.
Multiscale Models for the Mechanical Response of FCC Alloys under High Strain Rates and Complex Triaxial Loads [Aug 22 - Present]
Created an innovative multiscale model for predicting metallic alloys' response to high-strain-rate loads.
Merged large-scale atomistic simulations with machine learning for optimal model development.
Achieved improved results, capturing various deformation mechanisms under extreme conditions.
Liquid Time Constant Networks for Engineered Systems [Jan 22 - Aug 22]
Evaluated and demonstrated the superior performance of Liquid Time-Constant (LTC) networks in learning dynamics from noisy data compared to traditional recurrent neural networks.
Used synthetic, corrupted data to test the robustness of these networks under various conditions and parameter settings.
Showcased the effectiveness of LTC networks in modeling standard oscillatory systems under diverse test excitations.
Data Driven Modeling of Turbocharger Turbine using Koopman Operator [Aug 21 - Jan 22]
Developed a predictive model for transient and steady-state behavior of a turbocharger using the Koopman operator approach, significantly improving the understanding and control design of the system.
Applied Extended Dynamic Mode Decomposition to approximate the action of the Koopman operator, using experimental data from a Cummins heavy-duty diesel engine, outperforming existing nonlinear autoregressive models with exogenous inputs.
Utilized enhanced sensor data for more accurate modeling, addressing gaps in manufacturer-provided maps, especially in wide operating regions and incorporating heat transfer effects for more comprehensive and realistic modeling.
Reinforcement Learning based Control for Active Heave Compensation [Sep 20 - Aug 21]
Implemented Deep Deterministic Policy Gradient (DDPG) algorithm to capture the experience of the RL agent during training trails.
The simulation results demonstrated upto 10% better heave compensation performance of RL controller as compared to a tuned Proportional-Derivative Control.
The performance of the proposed method was compared with respect to heave compensation, offset tracking, disturbance rejection, and noise attenuation.
Data Driven Control for Active Heave Compensation [Apr 20 - Nov 20]
Performed the model identification of the winch model using long short term memory (LSTM) based recurrent neural network.
The neural network was trained using pairs of input-output data where the input was the opposite of the net heave time history of an KCS ship and output was the control input to winch placed on board the ship.
In addition, the ability of the LSTM network in handling the hard constraints of the swash angle was also analysed.
A Comparative Study of Different Active Heave Compensation Approaches [Aug 19 - Aug 20]
Analyzed and compared the performance of various control strategies in keeping a suspended payload regulated from a KCS container ship while the vessel is subjected to changing sea conditions.
The control strategies implemented were Proportional-Derivative (PD) control, Model Predictive control (MPC), Linear Quadratic Integral (LQI) compensator and Sliding Mode control (SMC).
Simulations were performed in MATLAB/SIMULINK environment for three scenarios: no disturbance or measurement noise, with disturbance but no measurement noise and with measurement noise but no disturbance.
