Ventilation systems are used in gas turbine packages to control the air temperature, to protect electrical instrumentation and auxiliary items installed inside the enclosure and to ensure a proper dilution of potentially dangerous gas leakages. These objectives are reached only if the ventilation flow is uniformly distributed in the whole volume of the package, providing a good air flow quality as prescribed by international codes such as ISO 21789. To evaluate the effectiveness of the ventilation design, numerical computations are performed for several purposes, one of which is the identification of poorly ventilated portions of the enclosure. In fact, it is essential to accurately detect the regions which are less ventilated, since they could be prone to the accumulation of an accidental fuel gas leak. There are different approaches to identify these portions, such as decay regression or inlet source analysis, that require unsteady simulations of the flow field inside the package. The present work discusses the implementation of a new methodology using machine learning and artificial neural networks (ANN) to detect the poorly ventilated regions where a gas cloud can accumulate. The concentration of fuel gas is estimated starting from a steady-state computation without running a more expensive unsteady computation. The entire process is built around an accurate training of the ANN using a proper set of simpler test-cases that have been identified to match the characteristics of the gas turbine enclosure. During the training phase accuracy and overfitting of the ANN were monitored to ensure robustness of the method. The procedure is then applied to a real case scenario and the results are presented in this paper highlighting the main advantages of this approach respect to a conventional use of CFD analysis. Computations of the flow fields are carried out using OpenFOAM with RANS and U-RANS approaches, while the ANN is developed and trained in Python.

Gas turbines usually are installed inside an enclosure, which is used as protection from the external environment and to provide an acoustic insulation. A ventilation system is required to control the temperature inside the enclosed volume and to dilute any potential gas leakage that may come from faulty pipes or flanges. The system has to be properly designed to avoid any unexpected explosion which would generate an overpressure not contained by enclosure walls. The most common approach to predict the effectiveness of the ventilation system requires to perform CFD analyses, which are very expensive in computational terms. A new approach has been proposed by authors, using machine learning and artificial neural networks (ANN) to identify the poorly ventilated zones. This methodology has been further developed, optimized and applied to a real gas turbine packages of new generation. In the present paper the authors will show the application of this procedure to the LM9000 package and the comparison with the results predicted using conventional CFD techniques. The tangible improvement introduced by this methodology is that the computational time is reduced from about three weeks with the common CFD approach to few minutes. The artificial neural network is developed in a Python environment that is applied during the CFX post-process phase of a steady state CFD simulation, providing results equivalent to unsteady CFD simulation. Besides the immediate benefits of this particular application, the suggested approach looks to be a great candidate to substitute the conventional and time-consuming CFD simulations with a fast post-processing algorithm that is able to learn and self-optimize as long as it is used.

One of the issues of handling large CFD datasets and process them to derive important design correlations is the limitation in automating the post-processing of data. Machine learning techniques, developed to process large unlabelled dataset, can play a key role on this subject. In this work an unsupervised approach to isolate different flow features inside a 2D cascade is proposed and validated. The approach relies on machine learning methods and in particular on Exploratory Data Analysis (EDA) and Principal Component Analysis for the pre-processing of the data and on K-means clustering for the post-processing. The K-means algorithm was trained on a Design of Experiments (DoE) of over 140 cases of 2D linear cascade configurations to identify the boundary layer on the profiles and the wake downstream. Validation resulted in a perfect capability of identifying the regions of interest. Then a possible exploitation of this method is presented, to compute pressure losses downstream of the cascade and train an artificial neural network to make a regression able to extend data to all the possible combinations of geometrical and operating parameters of the cascade. The same algorithm was applied to 3D flow cascades of profiles with sinusoidal leading edges to stress its extrapolation capability in case of flow regimes not present in the training DoE.

LES computations have limited applications in turbomachinery predictions because of the formidable amount of resources they require. Due to the exponential increase of requirements with Reynolds number, LES is usually limited to elements with moderate flow velocities and to investigate flows characterized by multiple length and time scales that overlaps. It is the case of combustion, aeroacoustics, unstable range of operations such as stalled conditions. In all these cases LES can provide unique insights on thermo-fluid-dynamics. The main drawback is that LES is not only very expensive, but also extremely sensitive to inflow conditions. A number of studies pointed out that slight differences in LES inflow conditions can result in a very different flow development and therefore different prediction of performance, noise and so on. It is so sensitive that comparison of computations with different inflow conditions and same arrangement for other parameters result in completely diverging results. There are two major solutions to this problem. First, is to use a cyclic inflow channel to generate a fully turbulent profile to feed to the main simulation. Second, the use of a synthetic turbulence model to generate analytically an unsteady turbulent profile. The drawback of the first approach is the fact that a fully developed inflow can be un-realistic for most turbomachinery applications. On the other hand, the second approach was proved not to be able to correctly reproduce the statistics of turbulence and therefore the synthetic inflows do not provide a real solution to the problem. In this paper we discuss a novel method to generate LES inflow conditions, based on adversarial machine learning. In particular we trained a generative adversarial network (GAN) to reproduce the inflow conditions of a channel flow. In this way, the generator of the GAN is trained to correctly reproduce an unsteady turbulent profile, while the discriminator is used during the training phase as an adversarial agent for the generator. During the validation of the method both the discriminator and the generator will be used to validate the proposed methodology.