pykitPIV: A framework for flexible and reproducible virtual training of machine learning models in optical velocimetry

In recent years, machine learning (ML) has become an increasingly integral part of experimental fluid dynamics. This evolution calls for new tools that can emulate real wind tunnel experiments and generate high-quality training data, as well as challenging testbeds for ML model development. Here, we introduce pykitPIV , a Python library that provides flexible and reproducible virtual environments for training ML algorithms in optical velocimetry. The synthetic datasets and environments generated by pykitPIV mimic those obtained from particle image velocimetry (PIV) or background-oriented Schlieren (BOS) experimental techniques. The library seamlessly integrates with various ML models, such as convolutional neural networks, variational methods, active learning, and reinforcement learning. It allows human users or the ML agents to flexibly configure parameters that typically emerge in experimental settings, such as seeding density, laser-sheet characteristics, camera exposure, particle loss, and experimental noise. In addition, pykitPIV enables an atlas of challenging synthetic velocity fields derived from analytic formulations (both compressible and incompressible), where particle drift and diffusion effects in stationary isotropic turbulence can be incorporated through a simplified Langevin model. Furthermore, pykitPIV enables ML agents to interact with virtual experiments, assimilate real experimental data, and train on a variety of tasks using diverse sensory cues and rewards. Ultimately, our library aims to accelerate real-time experimental inference, facilitate autonomous experimentation, and enable building robust models from experimental data, thereby supporting current trends in velocimetry.