Optimisation strategies using adaptive reduced order models in lithium-ion battery design

• Autores: Ruth Medeiros González
• Directores de la Tesis: Fernando Varas Mérida (dir. tes.) , María Higuera Torrón (codir. tes.)
• Lectura: En la Universidad Politécnica de Madrid ( España ) en 2025
• Idioma: inglés
• Tribunal Calificador de la Tesis: José Manuel Perales Perales (presid.) , Ángel Velázquez López (secret.) , Elena Beatriz Martín Ortega (voc.) , Eva Balsa-Canto (voc.) , Jerónimo Rodríguez García (voc.)
• Programa de doctorado: Programa de Doctorado en Ingeniería Aeroespacial por la Universidad Politécnica de Madrid
• Calificación: Sobresaliente Cum Laude
• Enlaces
  • Tesis en acceso abierto en: Archivo Digital UPM
• Resumen

  • This thesis presents different optimisation strategies designed to efficiently solve computationally expensive problems using time- and parameter-adaptive reduced order models (ROMs). The methods are applied to lithium-ion battery cell design, demonstrating their effectiveness in reducing computational costs. The techniques are general and applicable to a wide range of engineering problems involving high-cost numerical simulations.

    The proposed strategies build upon a time-adaptive model order reduction methodology developed for the Doyle-Fuller-Newman model. In particular, time- and parameter-adaptive ROMs combine two key techniques: a time-adaptive ROM, constructed on-the-fly by alternating with short full order model (FOM) intervals, and a parameter-adaptive strategy, which reuses information from previously evaluated parameter sets to reduce FOM time in subsequent simulations. These adaptive ROMs allow for fast and accurate evaluations of the objective functions and their derivatives, enabling their use within different optimisation algorithms.

    A novel methodology to compute the sensitivities, the derivatives of the solution fields with respect to the optimisation parameters, using time-adaptive ROMs is presented in Chapter 2. This approach computes sensitivities simultaneously alongside the integration of time-adaptive ROMs, using the solution basis to represent its sensitivities. By avoiding the need to solve the FOM sensitivity equations, this method leads to a substantial reduction in computational cost. The strategy is applied in Chapter 3 to optimise the porosity and thickness of electrodes in G-NMC811 and G-LFP cells, using time- and adaptive- ROMs for objective function evaluation and sensitivity computation. The results demonstrate both accuracy and robustness, achieving acceleration factors greater than 25 compared to FOM-based optimisation. In all the studied cases, the use of adaptive ROMs reduces the total optimisation cost to a level comparable to that of a single full simulation.

    Chapter 4 introduces a hybrid optimisation strategy that combines global and local methods. A genetic algorithm (GA) is used to explore the design space, while local optimisation refines the most promising solution candidates. A clustering-based criterion is used to identify convergence patterns in the GA populations and to reuse ROM data across initial population evaluations. By exploiting the information of previous objective function evaluations, the use of time- and parameter-adaptive ROMs allows reducing the initial FOM interval to solve only 1% of the simulation time, substantially improving overall computational efficiency, achieving an acceleration factor of approximately 7 with respect to FOM-based approaches.

    Finally, a continuation-based method for solving unconstrained bi-objective optimisation problems using time- and parameter- ROMs is proposed in Chapter 5. The algorithm initiates from a known point within the Pareto optimal set and advances along the Pareto manifold by moving to points that satisfy the optimality condition, necessitating efficient calculation of solution sensitivities. Compared to GAs, which often require tens of thousands of objective function evaluations, the proposed continuation approach needs only a few tens of evaluations. To further reduce computational cost, a time- and parameter-adaptive strategy is employed, using information from previous Pareto point evaluations to enrich ROMs. As a result, the proposed approach achieves a speed-up of three orders of magnitude while providing a more accurate representation of the Pareto manifold.

    In summary, the time- and parameter-adaptive ROM strategies developed in this work lead to substantial improvements in the computational efficiency of optimisation processes involving high-fidelity models. These methodologies offer strong potential to enable efficient design processes across multiple engineering disciplines.