Digital twins are revolutionizing industries by leveraging machine learning and data-driven models to create dynamic synchronized representations of physical systems. These virtual counterparts operate in real time, bridging the gap between the physical and digital worlds to simulate, predict, optimize, and control system behaviors, thereby enhancing the performance and efficiency of their physical analogs. Their transformative potential is particularly evident in manufacturing, where they contribute to precision metrology and quality assurance. This research explores the implementation of digital twins in experimental fluid mechanics, focusing on real-time data integration from in-line coherent imaging setups. By incorporating real-time sensor data or validated datasets, the goal is to develop dynamic models that accurately represent the behavior of the system under varying operating conditions.

The thesis emphasizes the use of advanced deep learning algorithms, including artificial neural networks (ANNs) and Vision Transformers (ViT), to create an end-to-end model for analyzing Particle Image Velocimetry (PIV) data. Convolutional neural network (CNNs) blocks, based on optical flow estimation techniques, are used to extract flow patterns by learning spatial features and correlations from PIV images. To refine predictions by capturing temporal dependencies and transient behaviors, iterative recurrent CNN blocks are integrated. However, these deep learning models, typically trained on synthetic datasets with reference results derived from analytical equations or high-fidelity numerical models, face challenges in robustness and generalization when applied to real-world industrial scenarios. To address this, experimental PIV datasets, focused on flow past a circular cylinder, were generated to evaluate the performance of RAFT-PIV (Recurrent All-Pairs Field Transforms), the state-of-the-art CNN model for optical flow estimation (Paper A). These datasets included variations in key parameters such as particle size, seeding density, and displacement ranges. The study also investigated the influence of image preprocessing on RAFT-PIV performance. The results showed that the root mean squared errors between the RAFT-PIV predictions and the reference data ranged from 0.5 to 2 pixels. Errors decreased with an increase in particle density or a reduction in the maximum particle displacement, while image pre-processing had a minimal impact on model performance.

Given the complexity of fluid flow phenomena, characterized by intricate structures across multiple scales, traditional ANNs using convolutional operations may capture local features but often miss long-range dependencies. To overcome this limitation, this research introduces Twins-PIVNet, a novel framework that replaces traditional convolution-based feature extractors with attention-based vision transformers (Paper B). This approach improves the model’s capacity to capture both global and local contexts, selectively focusing on relevant features. Twins-PIVNet outperforms most of the well-established PIV methods on both synthetic and experimental datasets, achieving a significant reduction in computational costs and inference compared to other self-attention-based models for PIV. Moreover, the model also excels in generalization and robustness, performing effectively on experimental data not included in the training set. These neural networks are capable of handling the nonlinear dynamics characteristic of fluid systems, significantly enhancing the predictive capabilities of digital twins. This advancement facilitates real-time analysis, predictive maintenance, and optimization, highlighting the critical role of digital twins in the advancement of flow analysis, smart manufacturing, and industrial optimization.