Physics-Enhanced Data-Driven Receding-Horizon Control for Agile Quadrotors and Electrochemical Systems

Abstract

Modern safety-critical systems are increasingly expected to operate near their physical limits while maintaining reliability, efficiency, and real-time implementability. In such settings, Model Predictive Control (MPC) and Moving Horizon Estimation (MHE) provide powerful optimization-based frameworks for control and state reconstruction, yet their practical effectiveness depends fundamentally on predictive models that are sufficiently accurate to capture performance and safety critical behavior while remaining computationally tractable for repeated online use. This challenge arises across the application domains considered in this dissertation. In agile quadrotor flight, reduced-order models neglect aerodynamic effects that become significant during aggressive maneuvers, whereas in lithium-ion and all-solid-state battery systems, reduced-order models may fail to represent internal degradation-relevant electrochemical states or coupled electro-mechanical behavior essential to safe operation. Motivated by this common modeling gap, this dissertation develops a physics-enhanced model-based framework for receding-horizon control and estimation that preserves the tractability of reduced-order models while selectively recovering the missing information most critical to performance and safety through learned residual augmentation, learned surrogate safety modeling, and explicit physics extension. An overview of the individual chapters is provided below. Chapter 3: This chapter develops a computationally efficient data-driven MPC method for agile quadrotor flight. Agile quadrotors operating at high speed experience substantial aerodynamic effects whose direct modeling is cumbersome and often computationally impractical for real-time optimization. Although combining Gaussian Process (GP) regression models with a simple dynamic model can significantly improve control performance, direct integration of GP models into the MPC pipeline introduces a considerable computational burden. To address this issue, this chapter presents an approach that separates GP inference from the online optimization by computing model corrections from the reference trajectory and current state measurements prior to solving the MPC problem. Validated in the Gazebo simulation environment, the proposed method demonstrates up to a 50% reduction in trajectory tracking error while matching the performance of direct GP integration with improved computational efficiency. Chapter 4: This chapter develops a data-based MHE method for agile quadrotors. Accurate state estimation is essential for precise trajectory control, yet the strong aerodynamic forces encountered during high-speed flight make this task particularly challenging. These complex turbulent effects are difficult to model, and the resulting unmodeled dynamics introduce inaccuracies into state estimation. To address this problem, Gaussian Processes are used to model the aerodynamic effects and are integrated into the MHE framework to achieve efficient and accurate state estimation with minimal additional computational burden. Through extensive simulation and experimental studies, the proposed method demonstrates substantial improvements in estimation performance and exhibits superior robustness in the presence of poor state measurements. Chapter 5: This chapter presents a unified data-driven framework for agile quadrotor flight that integrates GP-augmented MPC and GP-augmented MHE. Traditional control and estimation methods struggle with unmodeled aerodynamic disturbances, particularly during aggressive high-speed maneuvers. By learning these disturbances from prior data, the proposed GP-enhanced framework improves both control accuracy and estimation robustness within a single receding-horizon architecture. Simulation results in a ROS-Gazebo environment demonstrate up to 65% improvement in tracking performance together with significant gains in velocity estimation accuracy relative to baseline methods. Because the framework operates fully onboard using only GPS and IMU measurements, it is well suited for deployment in unstructured or visually degraded environments. These results demonstrate the potential of GP-augmented receding-horizon methods to enable more agile, reliable, and computationally efficient autonomous flight. Chapter 6: This chapter addresses safety-conscious fast charging of lithium-ion batteries, where charging speed is fundamentally limited by lithium plating, a degradation mechanism that compromises both safety and long-term performance. Whereas prior machine learning approaches in battery modeling have largely focused on voltage prediction, this chapter introduces a physics-enhanced data-driven charging framework that directly predicts plating risk and enforces safety constraints during charging. A GP surrogate is trained on high-fidelity Doyle-Fuller-Newman simulations to capture overpotential dynamics near the anode-separator interface, the region most susceptible to plating. This surrogate is embedded within MPC schemes based on the Single Particle Model with Electrolyte to impose physically meaningful safety constraints in real time. Simulation studies show that the GP-augmented MPC reduces lithium plating risk by up to 89, achieves nearly a tenfold reduction in peak overpotential violations, and decreases cumulative degradation by 95% relative to standard Constant-Current Constant-Voltage and nominal MPC strategies, with only marginal increase in total charging time. Furthermore, the same GP surrogate is integrated with an Equivalent Circuit Model, achieving comparable safety improvements with a 97% reduction in computational overhead, thereby supporting embedded implementation. Overall, this chapter establishes a modular, data-efficient, and computationally tractable pathway toward safer and faster charging of lithium-ion batteries. Chapter 7: This chapter addresses fast charging of all-solid-state batteries, whose promising energy and power density are accompanied by mechanical constraints arising from their solid-layer architecture. Solid-solid contacts and constrained volumetric expansion amplify internal stress generation, leading to cracking, interfacial delamination, and accelerated material degradation. Existing charging strategies generally neglect these deformation-driven limitations, while most available models either omit mechanics entirely or restrict them to simplified particle-level stress descriptions. To address this gap, this chapter develops a cell-level mechanical-electrochemical modeling and control framework that explicitly captures stress generation caused by volumetric expansion and its influence on allowable charging current. An equivalent Maxwell representation of the internal stress state is derived and coupled to an electrochemical model through a stress-dependent current limit that links compressive stress to the maximum admissible current density, such that operation within this limit suppresses harmful degradation. The resulting electro-mechanical model is embedded within MPC to regulate charging current and total strain rate while satisfying voltage, internal stress, and stress-dependent current constraints. Simulation results for an all-solid-state battery cell show that the proposed MPC can identify optimal charging profiles that suppress both mechanical and electrochemical degradation, while also serving as a tool to study the interaction between cell properties and charging policy design. Chapter 8: This chapter summarizes the main contributions of the dissertation and discusses future research directions. It highlights how physics-enhanced receding-horizon methods can improve control, estimation, and safety across distinct application domains by augmenting reduced-order models in ways that preserve real-time tractability.