Madhavan, Pramoth Varsan2026-05-262026-05-262026-05-262026-05-01https://hdl.handle.net/10012/23408The widespread deployment of proton exchange membrane (PEM) fuel cells requires the development of durable materials, optimized component architectures, and reliable system-level operation under realistic and often highly dynamic conditions. Progress across these domains is traditionally hindered by the extensive experimental effort, long testing times, and high costs associated with evaluating bipolar plate coatings, catalyst formulations, membrane electrode assembly (MEA) designs, and fuel cell behaviour under real-world automotive drive cycles. This thesis addresses these challenges by developing a multiscale, data-driven modelling framework that supports predictive insight and optimization at the materials, component, and systems levels. Through the use of machine learning, hybrid optimization algorithms, and advanced time-series models, this research aims to accelerate the design, performance assessment, and operational diagnostics of PEM fuel cells and supporting hydrogen infrastructure, contributing core elements toward future digital-twin-enabled systems. At the materials level, the thesis first investigates the corrosion resistance of metallic bipolar plates (MBPs), one of the most cost and durability critical components of PEM fuel cells. Stainless steel substrates coated with diamond-like-carbon material of varying thicknesses are experimentally evaluated using potentiodynamic polarization, electrochemical impedance spectroscopy, and surface wettability measurements. Using these datasets, extreme gradient boosting (XGB) and artificial neural network (ANN) models are developed to predict corrosion current density and impedance characteristics directly from coating thickness, electrochemical parameters, and contact angle values. The models achieved high predictive accuracy and successfully reproduced experimental trends, demonstrating that data-driven approaches can rapidly assess coating performance without requiring extensive physical testing. In parallel, the thesis addresses the optimization of oxygen reduction reaction (ORR) catalysts by developing an ANN-genetic algorithm (ANN-GA) framework capable of navigating the high-dimensional composition space of Pt-Co catalysts. Using experimental linear sweep voltammetry datasets obtained before and after accelerated stress tests, the hybrid framework identified catalyst compositions with improved catalytic activity and performance, demonstrating the potential of integrated machine learning and optimization tools to accelerate catalyst discovery. At the component level, the thesis developed a predictive and optimization framework for hybrid MEAs that combine catalyst-coated membrane (CCM) and catalyst-coated substrate (CCS) regions. This configuration has shown promise for enhancing water management, reactant transport, and overall electrochemical performance; however, determining the optimal CCM-to-CCS ratio remains a complex experimental challenge. An ANN-GA model is developed using single cell performance datasets under multiple operating conditions, including varied flow rates and backpressures. The framework accurately predicted polarization and power density curves and identified the optimal CCM_4_CCS_1 configuration (a hybrid MEA where the catalyst is distributed in a 4:1 ratio between the CCM and CCS sides), which outperformed both pure CCM and pure CCS structures. These results show that data-driven optimization can guide MEA design, reducing reliance on resource-intensive experimental screening. At the systems level, the thesis focuses on dynamic behaviour and diagnostic modelling. A long short-term memory (LSTM) network is constructed to predict transient thermal behaviour in a 50 cm2 PEM fuel cell under the new European driving cycle (NEDC). The model incorporated 72 input features, including 24 operating parameters and 48 spatial current distribution sensor readings, and generated 48 temperature predictions covering the 50 cm2 active area. Shapley additive explanations (SHAP) analysis revealed that both spatial current variations and operating conditions, such as reactant temperatures and pressures, strongly shape local thermal dynamics. The thesis also extends LSTM modelling to hydrogen infrastructure by predicting hydrogen valve outlet pressure across varied cycling conditions and temperatures (25°C, 85°C, and −40°C). The models achieved strong generalization and captured the long-term degradation signatures in valve behaviour, highlighting their suitability for predictive maintenance and operational safety. Collectively, the multiscale approaches developed in this thesis illustrate how data-driven models can provide rapid, accurate, and interpretable insights across materials, components, and system operation. Although independently developed, these methods collectively form a complementary set of predictive tools that support the broader vision of digital-twin PEM fuel cell systems. By advancing data-driven capabilities for materials screening, component optimization, and system-level diagnostics, this thesis lays essential groundwork for future platforms designed to enhance durability, enable real-time operational control, shorten development cycles, and support safe and efficient hydrogen energy technologies.enproton exchange membranefuel cellsmachine learningmultiscale optimizationbipolar platecatalystmembrane electrode assemblyhydrogen infrastructuredigital twinartificial intelligenceartificial neural networksgenetic algorithmextreme gradient boostinglong short-term memoryDoctoral Thesis