Ashwin Karichannavar

Monday, February 02, 2026, 10:30 - 11:30

Building 102 - SR 01-012

Abstract: Continuous aqueous two-phase flotation (ATPF) offers gentle biomolecule separation but is difficult to operate autonomously because of nonlinear, partially observed dynamics, model uncertainty, and stringent constraints on gas excitation. This thesis develops an offset-free nonlinear model predictive control (NMPC) architecture tailored to these challenges. A mechanistic process model and a black-box linear time-invariant conductivity model are identified from extensive ATPF datasets. An extended Kalman filter, augmented with integrating disturbance states, jointly estimates concentrations, conductivity, and steady-state biases. The associated noise covariances are identified from experimental data and prior sensor-noise characterization. Remaining tuning follows an innovation-based maximum-likelihood approach. A steady-state target selector computes disturbance-consistent references. A constrained NMPC—implemented in CasADi/IPOPT—optimizes inputs subject to magnitude, rate, and conductivity bounds.

Closed-loop simulations with a 3 minute sampling time and a 10-step prediction horizon validate the framework. Three scenarios are tested. The first scenario considers nominal conditions without noise. The second scenario injects measurement noise. The third scenario combines parametric model--plant mismatch on three perturbed plants with measurement noise and a mid-horizon reference step from 2 mol/m³ to 3 mol/m³. Across all cases, the controller achieves offset-free tracking of the outlet concentration, respects actuator and conductivity constraints, and maintains stable state estimates. Under mismatch, the estimated disturbance compensates model errors and removes steady-state offset with only minor overshoot. These results demonstrate that EKF based disturbance-state augmented estimation, steady-state target selection, and NMPC enables robust, constraint-satisfying operation of continuous ATPF. The framework also provides a foundation for future experimental validation.