Parameter-Dependent conditional Generative Adversarial Network (PDcGAN) Model for Multi-Phase Flow Prediction

Development of Parameter Dependent conditional Generative Adversarial Network (PDcGAN) Model for Multi-Phase Flow Prediction Machine learning and artificial intelligence (ML/AI) techniques have had a significant impact on various scientific and engineering fields by uncovering new insights from data or predicting previously untested properties. Deep learning, particularly in the form of artificial neutral networks, has been instrumental in the successes of ML/AI. Our research focuses on developing a deep learning framework using a conditional Generative Adversarial Network (cGAN) to predict the complex spray and air-fuel mixing in gasoline direct-injection (GDI) engines. This has been a major challenge due to the complicated two-phase flow dynamics involved. Our Parameter Dependent cGAN (PDcGAN) predicts the fuel’s 3D shape based solely on its sprayed condition and fuel property. Our training dataset consists of nine different fuels sprayed through multi-hole injectors 100 times each and projected into a combustion chamber. The projected images are acquired from three different camera angles and later used for 3D reconstruction as post-processing. These images are converted to pixel intensity data based on physically quantitative projected liquid volume (PLV) and separated into a training and validation set before training with the PDcGAN algorithm implemented in MATLAB. After extensive training on a GPU, the model can predict the morphology of fuels not included in the training data. Our research has identified optimal parameters and network architecture that yield an average test set error of around 10-15%. While this error rate may be considered high for practical use, it is important to note that engineers typically spend three hours using CFD (computational fluid dynamics) to predict the shape of a single fuel, requiring significant computational resources. In contrast, our trained algorithm, PDcGAN, provides results in seconds, making it a more efficient solution for engineers without access to supercomputing resources. We anticipate that the model’s accuracy will continue to improve as we expand the dataset to include a wider range of fuels.