In this work, a data-driven methodology for modeling combustion kinetics, Learned Intelligent Tabulation (LIT), is presented. LIT aims to accelerate the tabulation of combustion mechanisms via machine learning algorithms such as Deep Neural Networks (DNNs). The high-dimensional composition space is sampled from high-fidelity simulations covering a wide range of initial conditions to train these DNNs. The input data are clustered into subspaces, while each subspace is trained with a DNN regression model targeted to a particular part of the high-dimensional composition space. This localized approach has proven to be more tractable than having a global ANN regression model, which fails to generalize across various composition spaces. The clustering is performed using an unsupervised method, Self-Organizing Map (SOM), which automatically subdivides the space. A dense network comprised of fully connected layers is considered for the regression model, while the network hyper parameters are optimized using Bayesian optimization. A nonlinear transformation of the parameters is used to improve sensitivity to minor species and enhance the prediction of ignition delay. The LIT method is employed to model the chemistry kinetics of zero-dimensional and -air combustion. The data-driven method achieves good agreement with the benchmark method while being cheaper in terms of computational cost. LIT is naturally extensible to different combustion models such as flamelet and PDF transport models.
2 April 2019 in SAE International Journal of Advances and Current Practices in Mobility
Reduced-order models typically assume that the flow through the injector orifice is quasi-steady. The current study investigates to what extent this assumption is true and what factors may induce large-scale variations. Experimental data were collected from a single-hole metal injector with a smoothly converging hole and from a transparent facsimile. Gas, likely indicating cavitation, was observed in the nozzles. Surface roughness was a potential cause for the cavitation. Computations were employed using two engineering-level Computational Fluid Dynamics (CFD) codes that considered the possibility of cavitation. Neither computational model included these small surface features, and so did not predict internal cavitation. At steady state, it was found that initial conditions were of little consequence, even if they included bubbles within the sac. They however did modify the initial rate of injection by a few microseconds. Though the needle was never stationary, the mass discharge by the nozzle remained constant for most of the injection. The momentum discharge was more sensitive to lower needle lifts than the mass flow rate. An annular jet, that may follow either the needle surface or the sac wall, forms at low needle lift. The presence of this jet corresponds to a loss of momentum through the nozzle exit. The coefficient of area remains remarkably consistent during the early/late needle transient and is an important discovery.
Pipeline integrity is an important area of concern for the oil and gas, refining, chemical, hydrogen, carbon sequestration, and electric-power industries, due to the safety risks associated with pipeline failures. Regular monitoring, inspection, and maintenance of these facilities is therefore required for safe operation. Large standoff magnetometry (LSM) is a non-intrusive, passive magnetometer-based mea- surement technology that has shown promise in detecting defects (anomalies) in regions of elevated mechanical stresses. However, analyzing the noisy multi-sensor LSM data to clearly identify regions of anomalies is a significant challenge. This is mainly due to the high frequency of the data collection, mis-alignment between consecutive inspections and sensors, as well as the number of sensor measurements recorded. In this paper we present LSM defect identification approach based on ma- chine learning (ML). We show that this ML approach is able to successfully detect anomalous readings using a series of methods with increasing model complexity and capacity. The methods start from unsupervised learning with "point" methods and eventually increase complexity to supervised learning with sequence methods and multi-output predictions. We observe data leakage issues for some methods with randomized train/test splitting and resolve them by specific non-randomized splitting of training and validation data. We also achieve a 200x acceleration of support-vector classifier (SVC) method by porting computations from CPU to GPU leveraging the cuML RAPIDS AI library. For sequence methods, we develop a customized Convolutional Neural Network (CNN) architecture based on 1D convolutional filters to identify and characterize multiple properties of these defects. In the end, we report the scalability of the best-performing methods and compare them, for viability in field trials.
The adoption of Machine Learning (ML) for building emulators for complex physical processes has seen an exponential rise in the recent years. While neural networks are good function approximators, optimizing the hyper-parameters of the network to reach a global minimum is not trivial, and often needs human knowl- edge and expertise. In this light, automatic ML or autoML methods have gained large interest as they automate the process of network hyper-parameter tuning. In addition, Neural Architecture Search (NAS) has shown promising outcomes for improving model performance. While autoML methods have grown in popularity for image, text and other applications, their effectiveness for high-dimensional, complex scientific datasets remains to be investigated. In this work, a data driven emulator for turbulence closure terms in the context of Large Eddy Simulation (LES) models is trained using Artificial Neural Networks and an autoML frame- work based on Bayesian Optimization, incorporating priors to jointly optimize the hyper-parameters as well as conduct a full neural network architecture search to converge to a global minima, is proposed. Additionally, we compare the effect of using different network weight initializations and optimizers such as ADAM, SGDM and RMSProp, to explore the best performing setting. Weight and function space similarities during the optimization trajectory are investigated, and critical differences in the learning process evolution are noted and compared to theory. We observe ADAM optimizer and Glorot initialization consistently performs better, while RMSProp outperforms SGDM as the latter appears to have been stuck at a local minima. Therefore, this autoML BayesOpt framework provides a means to choose the best hyper-parameter settings for a given dataset.
It is well-accepted that the ability to go from one place to another, or mobility, contributes significantly to one's wellbeing. The need for mobility is universal, but the demand for mobility shows a great variation on a country basis. This particular study looks at what are some of the most important factors on a global level that can help in predicting the passengerkilometers-travelled or passenger-miles-travelled (PKT/PMT) on a country by country basis. This particular work tries to quantify the impact of some of the key variables like Gross Domestic Product (GDP), population growth, employment rate, number of households, age demographics within the population and macroeconomic variables on the total vehicle-based travel within each country. A panel-based regression model is developed to identify the effect of some of the key macroeconomic variables on the countries' PKT growth.
We propose a new deep learning based framework to identify pedestrians, and caution distracted drivers, in an effort to prevent the loss of life and property. This framework uses two Convolutional Neural Networks (CNN), one which detects pedestrians and the second which predicts the onset of drowsiness in a driver, is implemented on a Raspberry Pi 3 Model B+, shows great promise. The algorithm for implementing such a low-cost, low-compute model is presented and the results discussed.