Experimental results demonstrate that droplet vaporization on metal surfaces can be significantly enhanced with the application of a nanoporous, superhydrophilic surface coating. A thin layer of ZnO nanopillars can be easily seeded and grown on most metallic surfaces to achieve nanoscale pores between pillars, and ultra-low apparent contact angles. These surface coatings have the potential to improve spray cooling processes, and can be easily scaled up to larger and more complex heat exchangers. In order to characterize the potential improvement to a spray cooling system it is important to understand the dependence on system parameters, and to have a clear model of droplet vaporization on such surfaces. There are a number of surface and impact parameters that will affect the droplet spreading and subsequent vaporization on the surface. The surface contact angle, wicking speed and impact velocity all interact to affect the maximum spread of the droplet and the speed at which the droplet reaches this state. Along with variations in droplet volume and wall superheat, the model for droplet vaporization becomes more complex and nonlinear. Instead of exploring a single parameter at a time, machine learning tools can be utilized to determine the dependence of droplet evaporation time on these parameters simultaneously. In this study a genetic algorithm and a neural network were used to develop a droplet evaporation model for these superhydrophilic surfaces. Each algorithm demonstrated clear advantages depending on whether speed, accuracy, or an explicit mathematical model was prioritized.

Water temperature is an important parameter that influences the life near watercourses. The physical and chemical properties as the density and viscosity, gas solubility in water and in particular of the oxygen, the kinematics of the micro-biological and chemical phenomena affects not only the aquatic life but also the human life and animals living near the watercourse. In the Jiu river is discharged the water coming from the cooling towers of three thermo-electric power plants placed upstream. According to the recent researches realized by the EU Environmental Commission in Romania until 2050 the average temperature could rise by about 2°C. This paper presents the ability of a “neural networks” model to predict the water temperature in the lower part of the Jiu River. A database realized in over nine years in the measuring station Izbiceni, from the Olt County is used, concerning the water temperature (T w ) and the air temperature (T a ). The developed neural network is structured on five input variables and 9 hidden layers, 7 as sub-input and 2 as output. The stochastic time series is based on the separation of values for water T w and air T a in seasonal and daily fluctuations components. The schematic structure of the implemented neural network is presented, for obtaining the short-term and the seasonal components of the temperatures. The contact points are adjusted by the method of reprogramming after a gradient algorithm, in order to minimize the quadratic error between the output results and the relevant observed values. The performance of the neural network is based on estimation the thermal changes of the operation mode. For testing the numerical model, the calculations were performed using only the data recorded in nine years, in order to estimate a forecast for the next two years. The results are very close to the experimental recorded data and the obtained errors are less than 0.5°C. At the end of the paper are presented some results, conclusions and references.

Developing accurate and stable solutions for inverse heat conduction problems (IHCPs) is crucial in many industrial applications where direct measurement of surface conditions, such as heat flux or temperature, is not possible in practice and temperature measurement from interior points can be obtained alternatively. IHCPs are mathematically ill-posed and therefore developing stable solutions for them is challenging. Application of intelligent algorithms for solving IHCPs has been successfully explored for several cases. In the present paper, the problem of near real-time surface heat flux estimation in a one-dimensional domain with temperature dependent material properties and moving boundary is considered. An artificial neural network (ANN) is developed to use the temperature measurement data from interior points for limited number of time steps as the inputs and calculate the surface heat flux and recession rate at the current time step as the output of the network. For this purpose, a multi-layer perceptron (MLP) network is selected, trained and tested using heat flux-temperature data that were evaluated via COMSOL Multiphysics for a 1D medium that is exposed to standard heat flux profiles on its surface (including triangular, parabolic and step function). A randomly generated heat flux profile is also applied to the surface of the medium and temperature distribution is calculated via COMSOL Multiphysics. The temperature data are then used as the inputs to the network and surface heat flux is evaluated under this condition to assess the capability of the developed ANN in surface heat flux estimation. The performance of the network when using different number of inputs (previous and future time steps from which temperature data are needed for surface heat flux estimation) as well as different network topology are explored in the presence of random measurement error. The results show that the developed approach allows accurate near real-time surface heat flux estimation in a 1-D medium with temperature dependent material properties and moving boundary. The solution of this problem can be further extended to be used in sensors for ablative thermal protection system in space vehicles.

In recent times, goals for industry standards and national mandates have resulted in attempts to reduce the environmental impact of transient thermal processes (e.g., thermal management) in a multitude of applications ranging from industry to domestic use (consumer markets). A potential cheap, efficient and reliable solution is the implementation of a thermal energy storage (TES) unit which can serve as a primary or supplemental option, i.e., as a source of heating and/or cooling. Phase change materials (PCMs) can be used in TES due to their high latent heat storage capacity during phase transformation. Inorganic PCMs typically have the highest latent heat capacity and are attractive for their ability to store the highest amounts of thermal energy in small form factors while conferring respectable power ratings (however, they suffer from compromised reliability issues, that often arise from the need for subcooling). Subcooling (also known as supercooling) is a phenomenon where the temperature needs to be reduced substantially below the melting point to initiate solidification. A technique for obviating subcooling issues is to allow a small portion of the PCM to remain unmelted, which then allows the PCM to solidify and nucleate — starting from the unmelted portion of PCM (this is termed as the “cold finger” technique). A fundamental challenge for using the cold finger technique is to reliably control the amount of melt fraction in the total volume of the PCM (such that a target amount of the PCM remains solidified or unmelted). Thus, reliability is enhanced at the expense of substantial reduction in storage capacity. However, using Machine Learning (ML) techniques, this deficiency can be addressed by reliably predicting and thus controlling the amount of melt fraction in the total volume of the PCM with a higher accuracy than conventional techniques. Conventional techniques for predicting transient characteristics in real time control typically use physics-based prediction strategies that are effective for a narrow range of operating conditions with concomitant disadvantages: they suffer from high measurement uncertainties corresponding to large levels of error in the real time predictions and unreliability for a variety of operating conditions. In this study, significant technological advances were achieved by utilizing nearest neighbor search processes (such as radial basis functions) along with an artificial neural network (ANN) algorithm. This technique is simple to implement, device independent and was demonstrated successfully for predicting the melt fraction of a PCM with high accuracy and robustness. With this method, the melt fraction of a PCM can be accurately determined, which allows the maximum thermal capacity of a PCM to be utilized while mitigating reliability issues (such as subcooling) and enhancing the thermodynamic efficiencies of the TES platforms. Melting experiments were performed using a digital camera (for video recording) and a graduated cylinder containing PCM for monitoring the transient values of the melt fraction based on the height of the liquid phase of the PCM in the cylinder. An array of 3 thermocouples was placed at predetermined heights within the body of the PCM to monitor the transient propagation of the melting process within the PCM. In the final stages of the melting process, the predictions from the ML algorithm was found to be more accurate (90∼95% accuracy) than that of the conventional techniques based on physics-based solvers (∼60% accuracy). The accuracy of the ML algorithm was low at smaller melt fractions (∼30%) and improved substantially at higher melt fractions (∼95%).

The performance on the air side flow is often limited due to its lower heat transfer coefficient. This work is related to numerical simulation to study the significance of employing delta winglets in flat finned and wavy finned-tube heat exchangers. For this purpose, three-dimensional simulation data and a multi-objective genetic algorithm are employed. The angle of attack (α) of delta winglets and Reynolds number varied from 15° to 75° and 500 to 1300, respectively. Employing delta winglets has increased the heat transfer per unit temperature and per unit volume ( Z ) and the fan power per unit core volume ( E ) for both flat finned and wavy finned-tube heat exchangers. To achieve a maximum heat transfer enhancement and a minimum friction factor, the optimal values of these parameters ( Re and α) are calculated using the Pareto optimal strategy. For this purpose, CFD data, a surrogate model (neural network) and a multi-objective optimization genetic algorithm are combined. Results show that the performance of wavy finned-tube heat exchangers is higher than flat-finned tube heat exchangers which signify the importance of delta winglets in the wavy finned-tube heat exchangers.

The estimation of heat release rate (HRR) in a building fire is a meaningful yet challenging task to improve first emergency response. Inspired by the capability of deep learning method that mines patterns from raw data, the inverse model based on Gated Recurrent Unit (GRU) is presented for the inversion of HRR. First, a series of fire scenarios is simulated to form the dataset by forward fire modeling. Second, GRU is applied to learn representations from windows of time-series sensor data. Output layer is used to map the learned representations to targets. Third, the HRR is estimated by the trained GRU network with observed data from the fire. Finally, the accuracy and efficiency of the inverse model are evaluated in a multi-compartment configuration. Preliminary results indicate that GRU network can be applied to the inversion of fire HRR with higher accuracy and efficiency.

Heat transfer is a naturally occurring phenomenon which can be greatly enhanced by introducing longitudinal vortex generators (VGs). As the longitudinal vortices can potentially enhance heat transfer with small pressure loss penalty, VGs are widely used to enhance the heat transfer of flat-plate type heat exchangers. However, there are few researches which deal with its thermal optimization. Three dimensional numerical simulations are performed to study the effect of angle of attack and attach angle (angle between VG and wall) of vortex generator on the fluid flow and heat transfer characteristics of a flat-plate channel. The flow is assumed as steady state, incompressible and laminar within the range of studied Reynolds numbers ( Re = 380, 760, 1140). In the present work, the average and local Nusselt number and pressure drop are investigated for Rectangular vortex generator (RVG) with varying angle of attack and attach angle. The numerical results indicate that the heat transfer and pressure drop increases with increasing the angle of attack to a certain range and then decreases with increasing angle of attack. Moreover, the attach angle also plays an importance role; a 90° attach angle is not necessary for enhancing the heat transfer. Usually, heat transfer enhancement is achieved at the expense of pressure drop penalty. To find the optimal position of vortex generator to obtain maximum heat transfer and minimum pressure drop, the data obtained from numerical simulations are used to train a BRANN (Bayesian-regularized artificial neural network). This in turn is used to drive multi-objective genetic algorithm (MOGA) to find the optimal parameters of VGs in the form of Pareto front. The optimal values of these parameters are finally presented.

In this paper a genetic algorithm-polynomial neural network approach is used in order to model the effect of important parameters on heat transfer as well as fluid flow characteristics for a double-pipe helical heat exchanger by using numerical-certified results. In this way, overall heat transfer coefficient ( U o ), inner and annular pressure drop ( ΔP in , ΔP an ) are modeled with respect to the variation of inner and annular dean number, inner and annular Prandtl number, and pitch of coil which are defined as input (design) variables. The numerical-certified data was randomly divided into test and train sections which the former is used for benchmark. The GA-PNN structure was instructed by 75 percent of the numerical-validated data. 25 percent of the primary data which had been considered for testing procedure were entered into GA-PNN proposed models and results were compared by statistical criteria.

The earth is an energy resource which has more suitable and stable temperatures than air. Ground Source Heat Pumps (GSHPs) were developed to use ground energy for residential heating. The most important part of a GSHP is the Ground Heat Exchanger (GHE) that consists of pipes buried in the soil and is used for transferring heat between the soil and the heat exchanger of the GSHP. Soil composition, density, moisture and burial depth of pipes affect the size of a GHE. Design of GSHP systems in different regions of US and Europe is performed using data from an experimental model. However, there are many more techniques including some complex calculations for sizing GHEs. An experimental study was carried out to investigate heat transfer in soil. A three-layer network is used for predicting heat transfer from a buried pipe. Measured fluid inlet temperatures were used in the artificial neural network model and the fluid outlet temperatures were obtained. The number of the neurons in the hidden layer was determined by a trial and error process together with cross-validation of the experimental data taken from literature evaluating the performance of the network and standard sensitivity analysis. Also, the results of the trained network were compared with the numerical study.

Heat exchangers using in-tube condensation have great significance in the refrigeration, automotive and process industries. Effective heat exchangers have been rapidly developed due to the demand for more compact systems, higher energy efficiency, lower material costs and other economic incentives. Enhanced surfaces, displaced enhancement devices, swirl-flow devices and surface tension devices improve the heat transfer coefficients in these heat exchangers. This study is a critical review on the determination of the condensation heat transfer coefficient of pure refrigerants flowing in vertical and horizontal tubes. The authors’ previous publications on this issue, including the experimental, theoretical and numerical analyses are summarized here. The lengths of the vertical and horizontal test sections varied between 0.5 m and 4 m countercurrent flow double-tube heat exchangers with refrigerant flowing in the inner tube and cooling water flowing in the annulus. The measured data are compared to theoretical and numerical predictions based on the solution of the artificial intelligence methods and CFD analyses for the condensation process in the smooth and enhanced tubes. The theoretical solutions are related to the design of double tube heat exchangers in refrigeration, air conditioning and heat pump applications. Detailed information on the in-tube condensation studies of heat transfer coefficient in the literature is given. A genetic algorithm (GA), various artificial neural network models (ANN) such as multilayer perceptron (MLP), radial basis networks (RBFN), generalized regression neural network (GRNN), and adaptive neuro-fuzzy inference system (ANFIS), and various optimization techniques such as unconstrained nonlinear minimization algorithm-Nelder-Mead method (NM), non-linear least squares error method (NLS), and Ansys CFD program are used in the numerical solutions. It is shown that the convective heat transfer coefficient of laminar and turbulent condensing film flows can be predicted by means of theoretical and numerical analyses reasonably well if there is a sufficient amount of reliable experimental data. Regression analysis gave convincing correlations, and the most suitable coefficients of the proposed correlations are depicted as compatible with the large number of experimental data by means of the computational numerical methods.

Heat exchangers using in-tube condensation have great significance in the refrigeration, automotive and process industries. Effective heat exchangers have been rapidly developed due to the demand for more compact systems, higher energy efficiency, lower material costs and other economic incentives. Enhanced surfaces, displaced enhancement devices, swirl-flow devices and surface tension devices improve the heat transfer coefficients in these heat exchangers. This study is a critical review on the determination of the condensation heat transfer coefficient of pure refrigerants flowing in vertical and horizontal tubes. The authors’ previous publications on this issue, including the experimental, theoretical and numerical analyses are summarized here. The lengths of the vertical and horizontal test sections varied between 0.5 m and 4 m countercurrent flow double-tube heat exchangers with refrigerant flowing in the inner tube and cooling water flowing in the annulus. The measured data are compared to theoretical and numerical predictions based on the solution of the artificial intelligence methods and CFD analyses for the condensation process in the smooth and enhanced tubes. The theoretical solutions are related to the design of double tube heat exchangers in refrigeration, air conditioning and heat pump applications. Detailed information on the in-tube condensation studies of heat transfer coefficient in the literature is given. A genetic algorithm (GA), various artificial neural network models (ANN) such as multilayer perceptron (MLP), radial basis networks (RBFN), generalized regression neural network (GRNN), and adaptive neuro-fuzzy inference system (ANFIS), and various optimization techniques such as unconstrained nonlinear minimization algorithm-Nelder-Mead method (NM), non-linear least squares error method (NLS), and Ansys CFD program are used in the numerical solutions. It is shown that the convective heat transfer coefficient of laminar and turbulent condensing film flows can be predicted by means of theoretical and numerical analyses reasonably well if there is a sufficient amount of reliable experimental data. Regression analysis gave convincing correlations, and the most suitable coefficients of the proposed correlations are depicted as compatible with the large number of experimental data by means of the computational numerical methods.

We introduce a methodology to extract the regimes of operation from condensing heat exchanger data. The methodology uses a Gaussian mixture clustering algorithm to determine the number of groups from the data, and a maximum likelihood decision rule to classify the data into these clusters. In order to assess the accuracy of clustering technique, experimental data from the literature visually classified as dry-surface, dropwise condensation, and film condensation, are used in the analysis. Though there is a discrepancy between the clustering classification and the visual one, an independent evaluation using artificial neural networks (ANNs) shows that the clustering methodology is able to both find the different regimes of operation and classify the data corresponding to each regime.

Artificial neural network (ANN) has shown its superior predictive power compared to the conventional approaches in many studies. However, it has always been treated as a “black box” because it provides little explanation on the relative influence of the independent variables in the prediction process. In our previous work (Tam et al., 2006), an index of contribution extracted from the ANN correlation was primarily introduced to analyze the relative importance of the associated independent variables on our forced convective turbulent heat transfer data in a horizontal tube (Ghajar and Tam, 1994). The most and the least important variables were determined quantitatively and found to be thoroughly conforming to the empirical correlation and physical phenomena. In this study, we have extended the method to a more complicated data set, forced and mixed convection developing laminar flow in a horizontal tube with uniform wall heat flux. The parameters influencing the Nusselt number for this data set were Reynolds number, Grashof number, Prandtl number, the length-to-diameter ratio, and the bulk-to-wall viscosity ratio. Due to the complexity of the problem it is difficult to determine the influence of the individual independent variables. According to literature, for laminar heat transfer involving entrance and mixed convection effects, Rayleigh number and Graetz number are both important. Through the re-arrangement of those variables, the factor analysis clearly showed that the Rayleigh number has a significant influence on the mixed convection heat transfer data and the forced convection heat transfer data is more influenced by the Graetz number. The results clearly indicate that the factor analysis method can be used to provide an insight into the influence of different variables or a combination of them on complicated heat transfer problems.

Turbulent forced convection correlations are available in the literature for gases (Pr ∼ 0.7), but the test data leave a gap in the range of Prandtl (Pr) number between 0.1 and 0.7 occupied by binary gas mixtures. In this paper we develop a turbulent forced convection correlation for the Nusselt (Nu) number of in-tube binary gas mixtures for the ranges of Reynolds (Re) number between 10 4 and 10 6 and Prandtl (Pr) number between 0.1 and 0.7. A fully connected back-propagation Artificial Neural Network (ANN) is used to learn the pattern of Nu as a function of Re and Pr. Available test data in the range of 0.001 < Pr < 0.1 and 0.7 < Pr < 1000 are provided to the ANN. The test data are separated in two sets to train and test the neural network. A training set with 80% of the data is used to predict a testing set with the remaining 20% of the data. After the network is trained, we make use of the excellent nonlinear interpolation capabilities of ANNs, to predict values of Nu for the sought range 0.1 < Pr < 0.7. These predictions are later used to generate a correlation that aptly covers the complete range of Prandtl numbers.

This work uses an artificial neural network (ANN) to correlate experimentally determined Nusselt numbers and friction factors for three kinds of fin-and-tube heat exchangers having plain fin, slit fin and fin with longitudinal winglet vortex generators with large tube-diameter and large tube-row. First, the relatively limited experimental data was picked up from the database of nine samples with Reynolds number being between 4,000 and 10,000. The Back Propagation (BP) algorithm was used to train the networks. Compared with correlations for prediction using conventional power-law regressions, the performance of developed ANN based prediction exhibits ANN superiority. Different network configurations were assessed to find the best architecture for correlating heat transfer and flow friction. The deviation between the predictions and experimental data was less than 4%. Then the ANN training database was expanded to include the experimental data and numerical data of some similar geometries by Computational Fluid Dynamics (CFD), which in turn indicated that the predictions agree well with the combined database. The satisfactory results suggest that the ANN model is generalized to correlate the heat transfer and fluid flow of such three kinds of heat exchangers with large tube-diameter and large tube-row. It is recommended that ANNs might be used to predict the performance of thermal systems in engineering applications.

In this paper the Support Vector Machines (SVM) method is used to correlate the transitional forced and mixed convection experimental data of Ghajar and Tam (1994) that were obtained along a stainless steel horizontal circular tube fitted with re-entrant, square-edged, and bell-mouth inlets under uniform wall heat flux boundary condition. The SVM method has been chosen to further improve the accuracy of the correlations that were developed by Ghajar and his co-workers using the traditional least-squares method (Ghajar and Tam, 1994) and more recently the artificial neural networks (ANN) method (Ghajar et al., 2004). Using the ANN method improved the accuracy of their correlation. However, there are drawbacks associated with ANN method. One of the major problems with the ANN method is that it does not provide a unique correlation due to different coefficient matrices. The SVM method used in this study eliminated the drawbacks associated with the ANN method and provided a unique correlation with comparable accuracy as the ANN method. For the experimental data used, majority of the data points were predicted within 5% deviation. Comparisons were made regarding the accuracy of the developed correlation and its characteristic using SVM and ANN methods. The results showed that SVM is a good method to correlate complex heat transfer data.

In this paper weld pool shapes of type 304 stainless steel obtained through computational fluid dynamics simulations (CFD) will be used to train artificial neural networks. The neural network is then used to determine the experimental parameters of the ND:YAG laser system on 304 steel by feeding it with the parametrized, experimentally obtained weld pool parameters. The output of the neural network are the process parameters like laser spot diameter and laser power on the surface of the weld pool. These parameters are very difficult to determine accurately through experiments but are necessary inputs for CFD simulations of the weld pool. Accurate CFD simulation is required for simulation and modeling of the solidified micro structure of the material.

An inverse heat conduction problem (IHCP) for nanoscale structures was studied. The conduction phenomenon is modeled using the Boltzmann transfer equation. Phonon-mediated heat conduction in one dimension is considered. One boundary, where temperature observation takes place, is subjected to a known boundary condition and the other boundary is exposed to an unknown temperature. The artificial neural network (ANN) is employed to solve the described inverse problem. Sample results are presented and discussed.

The feasibility of a transportable artificial neural network (ANN)–based technique for the classification of flow regimes in three phase gas/liquid/pulp fiber systems, using pressure signals as input, was demonstrated in this study. Both supervised and unsupervised neural network models were applied for implementing regime classification. Data obtained in a vertical column (1.8m high and 5.08cm in diameter) were used, and a supervised ANN was designed and successfully tested that used some characteristics of the power density spectrum of the recorded signals of a pressure sensor as input. The developed ANN showed encouraging transportability. An ANN-based method was also developed for adjusting the processed signals of one sensor before feeding them as input to an ANN that had been trained based on data from another similar sensor. The method further improved transportability. The objectivity of the experimentally-identified flow regimes and their transition conditions was verified by the application of a Kohonen self-organizing neural network.

To achieve maximum hydraulic efficiency at a wide range of operation, efficient numerical tools are inevitable for the shape optimisation of hydraulic machinery, e.g. rotodynamic pumps and water turbines. A suitable geometry representation by means of B-Spline techniques for the three-dimensionally (3D) curved bladings is necessary as well as a powerful CFD technique in order to evaluate and analyze the hydraulic properties of the design. Both of these modules are integrated in one program system so as to allow the design engineer to achieve maximum performance. There is a natural contradiction between the numerical effort to describe a complex geometry and to compute the flow field as exactly as necessary for an accurate evaluation, and the industrial need to lower the costs and the duration of development of a new design. The 3D curved bladings of rotodynamic pumps and water turbines are characterized by a big number of parameters. Thus, the shape optimisation of this type of blades is still a problem to be solved by an efficient approach. To simplify this complexity, the blade profiles are defined by a reduced set of B-Spline describers on a number of stream surfaces in the conformal mapping representation. This provides an efficient and quick method for modifying the blade shape. To reduce the numerical effort in an iterative optimisation process, a new idea of using the Multi Level CFD-Technique (MLCFD) is described more detailed. The CFD tools used in this process, can be chosen by different criterions such as accuracy on the one hand and computation time on the other hand: There are quasi-3D and full-3D Euler codes as well as quasi-3D and full-3D Navier-Stokes codes for the flow computation and analysis. Starting from an initial design which may come from an artificial neural network having been trained from an existing database, a rough estimation of the quality of a geometry modification can be performed by a quick quasi-3D Euler computation, whereas the final evaluation of the optimum design should be carried out by an accurate 3D Navier-Stokes code. The MLCFD-technique is applied to optimize the blade shape of a centrifugal pump impeller with a specific speed n q = 39 l/min. The numerical results show that significant improvements of the local and integral flow quantities as well as a remarkable reduction of the numerical effort may be reached by using this new approach.