Deep learning-based optical metrology for real-time sensing: optical flow estimation

Advancements in deep learning have significantly broadened the scope of optical metrology, enabling real-time, high-precision sensing for industrial and environmental applications. This work investigates the integration of deep learning with optical flow estimation—specifically in particle image velocimetry (PIV), a non-intrusive flow visualization technique—to improve accuracy, computational efficiency, and adaptability. We introduce end-to-end models that reliably estimate dense flow fields from consecutive time-resolved particle images, optimized for resource-constrained environments. By utilizing core neural network operations such as automatic differentiation and backpropagation, the models extract higher-order flow moments, which are incorporated into training implicitly via physics-informed loss functions or explicitly as evaluation metrics. This hybrid strategy ensures physically consistent and interpretable predictions. Beyond PIV, the framework is adaptable to a wide range of optical metrology methods, offering a scalable solution across multiple domains. Its real-time capabilities position it as a valuable tool within digital twin frameworks, enabling predictive maintenance, adaptive control, and sustainable manufacturing through process optimization and surface quality assessment. By combining data-driven learning with governing physics, this approach bridges fluid dynamics, environmental monitoring, and advanced manufacturing—fostering interdisciplinary collaboration and enabling efficient, informed decision-making.