Employing a large language model and a Transformer model for data analysis of turbulent bubble flow

Large language models (LLMs) hold significant application value in the field of natural language processing and have demonstrated considerable potential in processing time series data. In hot-film measurements of bubble flow, bubble-induced interference with the hot-film data can affect the accuracy of turbulent flow field analysis. To address this issue, this study proposes a composite method integrating LLM-based identification of bubble interference signals and Transformer-based signal restoration. The core of this approach involves converting signals into natural language representations and employing a prompt as a signal prefix, enabling a time series-large language model (Time-LLM) to accurately recognize bubble-induced signal segments. Subsequently, the identified bubble signals are removed, and the resulting gaps are restored using a Transformer model. The results indicate that the proposed Time-LLM method can accurately identify bubble interference signals in hot-film measurements, and the Transformer model effectively restores the regions affected by bubble interference. Based on energy spectrum analysis, it is found that the turbulent scaling exponent in bubble-laden flows follows a −3 power law, which deviates from the classical −5/3 scaling law observed in isotropic turbulence.