Artificial intelligence (AI) is reshaping the landscape of experimental fluid mechanics by transforming how flow information is extracted, interpreted, and modeled. Despite its potential, adoption remains cautious, constrained by the perception that Artificial Neural Networks (ANNs) operate as opaque 'black boxes' lacking explicit physical grounding. This dissertation addresses that skepticism by demonstrating that, when purposefully designed, adapted, and constrained by governing equations, Deep Learning can function as a transparent and reliable extension of classical measurement science, capable of reconstructing temporally consistent and physically meaningful results directly from raw measurements. As experimental and computational capabilities expand, fluid mechanics now operate in a regime where the volume, resolution, and dimensionality of data far exceed what classical analysis frameworks were designed to handle. Extracting coherent flow information from such datasets requires methods capable of modeling nonlinear image dynamics, capturing long-range spatial interactions, and integrating temporal evolution. Deep Learning, with its inherent quality of hierarchical representation learning, offers an alternative paradigm: one that learns motion representations directly from raw experimental data. However, its successful integration into measurement science requires rigorous benchmarking, explicit physical grounding, and architectural innovation. This work advances the state-of-the-art in Particle Image Velocimetry (PIV) through the development of Deep Learning-based architectures for optical flow estimation, that couple spatial attention, temporal reasoning, and physics-informed learning.

The thesis begins with a systematic experimental evaluation of recurrent optical-flow estimation neural networks on canonical cylinder-wake datasets, quantifying how performance of these networks varies with seeding density, particle size, and flow regime (Paper A). These analyses reveal that while well-established Deep Learning-based models for PIV, such as RAFT-PIV (Recurrent All-Pairs Field Transforms) successfully recover dominant flow structures, their accuracy degrades in regions of strong shear, separation, or low signal-to-noise ratio, highlighting the need for architectures that reason globally and adapt to experimental variability. To address these limitations, a Vision Transformer (ViT)-based architecture, Twins-PIVNet, is introduced to leverage spatial self-attention for capturing multi-scale particle motion beyond the reach of conventional convolutional networks (Paper B). By learning global contextual relationships directly from raw particle images, Twins-PIVNet improves robustness and accuracy across a broad range of flow regimes and achieves sub-second inference, demonstrating the effectiveness of attention-based feature extraction for handling real-world PIV complexity.

Temporal coherence, an essential yet often overlooked attribute of experimental velocimetry, is then addressed through TriP-Net model, a multi-frame architecture that departs from the traditional two-frame paradigm used in most Deep Learning-based PIV (Paper C). By incorporating three consecutive particle images, TriP-Net captures higher-order temporal dynamics, yielding velocity fields with smoother evolution and reduced continuity errors. Building on this foundation, the framework is further extended into ADHD-PIV, a physics-informed architecture trained on four-frame sequences (Paper D). By coupling self-attention with recurrent updates and embedding adaptively weighted divergence and acceleration constraints, ADHD-PIV produces flow fields that better respect conservation laws and maintain stability over extended temporal windows. Together, these contributions move Deep Learning-based PIV from purely spatial inference toward fully temporal, physics-constrained flow reconstruction.

Collectively, these developments establish a unified and scalable methodology for Deep Learning in experimental fluid mechanics, one that is data-efficient, interpretable, and firmly grounded in physical principles. The models presented in this thesis reconstruct flow fields with sub-pixel precision and enhanced temporal coherence, enabling real-time analysis suitable for intelligent Digital Twins and advanced optical metrology (Paper E). By integrating high-fidelity experiments with self-supervised and physics-aware learning strategies, the work demonstrates how AI can evolve from a post-processing tool into a core component of measurement science, supporting real-time sensing, predictive diagnostics, and adaptive control. Ultimately, this thesis contributes both conceptually and practically to the integration of Artificial Intelligence in fluid mechanics, showing that trust in artificial neural networks arises not from revealing their internal algebra but from embedding the governing laws of fluid motion within them. The findings redefine how experimental flow measurements are performed and offer a pathway toward autonomous diagnostics and adaptive Digital Twin systems for complex thermal–fluid environments.