publications
An up-to-date list is available on Google Scholar.
2022
- Front. Phys.Enabling Real-Time Adaptation of Machine Learning Models at X-Ray Free Electron Laser Facilities with High-Speed Training Optimized Computational HardwarePetro Junior Milan, Hongqian Rong, Craig Michaud, Naoufal Layad, and 2 more authorsFrontiers in Physics, 2022
The emergence of novel computational hardware is enabling a new paradigm for rapid machine learning model training. For the Department of Energy’s major research facilities, this developing technology will enable a highly adaptive approach to experimental sciences. In this manuscript we present the per-epoch and end-to-end training times for an example of a streaming diagnostic that is planned for the upcoming high-repetition rate x-ray Free Electron Laser, the Linac Coherent Light Source-II. We explore the parameter space of batch size and data parallel training across multiple Graphics Processing Units and Reconfigurable Dataflow Units. We show the landscape of training times with a goal of full model retraining in under 15 min. Although a full from scratch retraining of a model may not be required in all cases, we nevertheless present an example of the application of emerging computational hardware for adapting machine learning models to changing environments in real-time, during streaming data acquisition, at the rates expected for the data fire hoses of accelerator-based user facilities.
2021
- JCPDeep-Learning Accelerated Calculation of Real-Fluid Properties in Numerical Simulation of Complex FlowfieldsPetro Junior Milan, Jean-Pierre Hickey, Xingjian Wang, and Vigor YangJournal of Computational Physics, 2021
A deep-learning based approach is developed for efficient evaluation of thermophysical properties in numerical simulation of complex real-fluid flows. The work enables a significant improvement of computational efficiency by replacing direct calculation of the equation of state with a deep feedforward neural network with appropriate boundary information (DFNN-BC). The proposed method can be coupled to a flow solver in a robust manner. Depending on the numerical formulation of the flow solver, the neural network takes in either the primitive or conservative variables, including the chemical composition of the system, and calculates all relevant fluid properties for the subsequent routines in the solver. Two test problems are employed to validate the proposed methodology. The first uses a preconditioning scheme with dual-time integration for the simulation of swirl rocket injector flow dynamics under supercritical conditions. The second uses a conservative-variable based formulation for the simulation of laminar counterflow diffusion flames for cryogenic combustion. A parametric analysis is performed to optimize the numbers of hidden layers and neurons per hidden layer. The computational accuracy, efficiency, and memory requirements of the neural network are examined. The DFNN-BC model accelerates the evaluation of real-fluid properties by a factor of 2.43 and 3.7 for the two test problems, respectively, and the overall flowfield simulation by 1.5 and 2.3, respectively. In addition, the memory usage is reduced by up to five orders of magnitude in comparison with the table look-up method.
- SAEAccelerating the Generation of Static Coupling Injection Maps Using a Data-Driven EmulatorSudeepta Mondal, Roberto Torelli, Bethany Lush, Petro Junior Milan, and 1 more authorSAE International Journal of Advances and Current Practices in Mobility, 2021
Accurate modeling of the internal flow and spray characteristics in fuel injectors is a critical aspect of direct injection engine design. However, such high-fidelity computational fluid dynamics (CFD) models are often computationally expensive due to the requirement of resolving fine temporal and spatial scales. This paper addresses the computational bottleneck issue by proposing a machine learning-based emulator framework, which learns efficient surrogate models for spatiotemporal flow distributions relevant for static coupling injection maps, namely total void fraction, velocity, and mass, within a design space of interest. Different design points involving variations of needle lift, fuel viscosity, and level of non-condensable gas in the fuel were explored in this study. An interpretable Bayesian learning strategy was employed to understand the effect of the design parameters on the void fraction fields at the exit of the injector orifice. The results show a strong influence of the amount of non-condensable gas on the level of cavitation as well as the overall shape of the gas-phase structures at the orifice exit. The emulator framework involves the construction of deep autoencoders for efficient dimensionality reduction of the flowfields. Deep artificial neural networks were then employed for prediction of the flowfields for unknown operating conditions. The emulated flowfields were then tested by evaluating spray and combustion predictions from one-way coupling spray simulations. The analysis of the spray predictions from CFD-generated and emulator-predicted injections maps revealed that the emulation framework is capable of reproducing spray predictions with similar level of accuracy, yet at a fraction of the computational cost. The maximum achievable speed-up using the emulator framework is up to 2 million times over the traditional CFD approach for generating static coupling injection maps. The emulation framework provides an efficient pathway for integrating detailed injector simulations into spray and engine simulations.
2020
- A&SData-Driven Model Reduction of Multiphase Flow in a Single-Hole Automotive InjectorPetro Junior Milan, Roberto Torelli, Bethany Lusch, and Gina M. MagnottiAtomization and Sprays, 2020