Distributed Bayesian optimization of deep reinforcement learning algorithms
Not Available
SEARCH · Engineering Papers
Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.
Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.
Not Available
Gasification is an important method of converting coal into clean burning fuels and high-value industrial chemicals. However, gasifier reliability can be severely limited by rapid degradation of the refractory lining in hot-wall gasifiers. The Pacific Northwest National Laboratory (PNNL) is performing multidisciplinary research to provide the experimental data and the engineering models needed to control gasifier operation for extended refractory life. Our experimental program includes prediction of slag viscosity using empirical viscosity models encompassing US coals, characterization of selected slag-refractory interaction including transport of slag/refractory components at the slag-refractory interface, and measurement of slag penetration into refractories as a function of time and temperature. The experimental data is used in slag flow, slag penetration, and refractory damage models to predict the operating temperature limits for increased refractory life. A simplified entrained flow gasifier model is also being developed to simulate one-dimensional axial flow with average axial velocity, coal devolatilization, and combustion kinetics. Combining the slag flow, refractory degradation, and gasifier models will provide a powerful tool to predict the coal and oxidant feed rates and control the gasifier operation to balance coal conversion efficiency with increased refractory life. A research scale gasifier has also been constructed at PNNL to provide syngas for coal conversion and carbon sequestration research, and also valuable datasets on operating conditions for validating the modeling results.
We present a brief survey of investigations carried out in recent years by the Lviv scientific school of the mechanics of coupled fields on the problems of construction and generalization of mathematical models and methods aimed at the description, determination, and optimization of the thermomechanical behavior of bodies under the combined action of force, thermal, and electromagnetic loads.
We consider a two-point boundary value problem (TPBVP) in orbital mechanics involving a small body (e.g., a spacecraft or asteroid) and N larger bodies. The least action principle TPBVP formulation is converted into an initial value problem via the addition of an appropriate terminal cost to the action functional. The latter formulation is used to obtain a fundamental solution, which may be used to solve the TPBVP for a variety of boundary conditions within a certain class. In particular, the method of convex duality allows one to interpret the least action principle as a differential game, where an opposing player maximizes over an indexed set of quadratics to yield the gravitational potential. In the case where the time duration is less than a specific bound, there exists a unique critical point for the resulting differential game, which yields the fundamental solution given in terms of the solutions of associated Riccati equations.
The goal of this paper is to solve a class of stochastic optimal control problems numerically, in which the state process is governed by an Itô type stochastic differential equation with control process entering both in the drift and the diffusion, and is observed partially. The optimal control of feedback form is determined based on the available observational data. In this work, we call this type of control problems the data driven feedback control. The computational framework that we introduce to solve such type of problems aims to find the best estimate for the optimal control as a conditional expectation given the observational information. To make our method feasible in providing timely feedback to the controlled system from data, we develop an efficient stochastic optimization algorithm to implement our computational framework.
Here, we study the relationship between the Quantum Approximate Optimization Algorithm (QAOA) and the underlying symmetries of the objective function to be optimized. Our approach formalizes the connection between quantum symmetry properties of the QAOA dynamics and the group of classical symmetries of the objective function. The connection is general and includes but is not limited to problems defined on graphs. We show a series of results exploring the connection and highlight examples of hard problem classes where a nontrivial symmetry subgroup can be obtained efficiently. In particular, we show how classical objective function symmetries lead to invariant measurement outcome probabilities across states connected by such symmetries, independent of the choice of algorithm parameters or number of layers. To illustrate the power of the developed connection, we apply machine learning techniques toward predicting QAOA performance based on symmetry considerations. We provide numerical evidence that a small set of graph symmetry properties suffices to predict the minimum QAOA depth required to achieve a target approximation ratio on the MaxCut problem, in a practically important setting where QAOA parameter schedules are constrained to be linear and hence easier to optimize.
Optimization of charge burden for foundry furnaces lies in determining a percentage or mass (kg, t) of the share of each charge material in the burden with the lowest unit cost. A classic task of charge burden optimization requires a definition of an objective function, which most often defines the cost of the charge burden and a system of constraints, including technological assumptions such as the balance of chemical elements for the required chemical composition of the charge or molten metal, limitations of the mass fraction of the individual elements of the charge, and others. The solution for such an optimization task is a set of starting values for the devices which weigh and proportion the charge materials for a foundry furnace. The consequences of deviations from proper values, causes by inaccurate weighing or proportioning and human error as well, can be limited by using an appropriately designed and implemented multistage optimization algorithm. This article presents a mathematical model of such an algorithm and discussion of its practical application. The algorithm as described below can be an important element of computer-aided decision to be made by employees in the process of batch compilation or automatically control devices dosing individual batch components. An IT system with an implemented algorithm may significantly increase the efficiency of charge preparation for a foundry furnace, guaranteeing its correct chemical composition despite batch materials dosing errors. It will contribute to a significant improvement of economic indicators of liquid metal production, with the effect of boosting the competitiveness of the foundry industry.
The SO{sub 2} sorption capacity (SSC) of sorbents prepared from rice husk ash (RHA) with NaOH as additive was studied in a fixed-bed reactor. Rice husk ash is produced by burning rice husks at about 300{sup o}C and was chosen as a source of siliceous material, abundantly available in Malaysia. The sorbents were prepared using a water hydration method by slurrying RHA, CaO, and NaOH. Response surface methodology (RSM) based on four-variable central composite face centered design (CCFCD) was employed in the synthesis of the sorbents. The correlation between the sorbent SSC (as response) with four independent sorbent preparation variables, i.e. hydration period, RHA/CaO ratio, NaOH amount, and drying temperature, were presented as empirical mathematical models. Among all the variables studied, the amount of NaOH used was found to be the most significant variable affecting the SSC of the sorbents prepared. The SSC for sorbent prepared with the addition of NaOH was found to be significantly higher than sorbents prepared without NaOH. This is probably because NaOH is a deliquescent material, and its existence increases the amount of water collected on the surface of the sorbent, a condition required for sorbent-SO{sub 2} reaction to occur at low temperature. The effect of further treatment of RHA at 600{sup o}C was also investigated. Although pretreated RHA sorbents demonstrated higher SSC as compared to untreated RHA sorbents, nevertheless, at optimum conditions, sorbents prepared from untreated RHA was found to be more favorable due to practical and economic concerns. 33 refs., 3 figs., 1 tab.
All solid-state lithium batteries are, potentially, higher energy density and safer alternatives to conventional lithium-ion batteries (LIBs). These are particularly attractive characteristics for large-scale applications such as electric vehicles and grid energy storage systems. However, the thin film deposition techniques used to make current devices are not readily scalable, and result in low areal capacities, which translate to low practical energy densities. To overcome these deficiencies, it is necessary to design thicker electrodes similar to what are used in LIBs (30–100 μm), in which the active material is composited with an ionic conductor and an electronically conducting additive, to overcome transport limitations. In this paper, we propose a method for making such an electrode, starting with a porous scaffold, i.e. Li 7 La 3 Zr 2 O 12 (LLZO), made by freeze casting, which is then infiltrated with the active material LiNi 0.6 Mn 0.2 Co 0.2 O 2 (NMC-622) and other components. The freeze casting technique results in the formation of oriented channels with low tortuosity, which run roughly parallel to the direction of the current. The scaffolds were characterized with synchrotron X-ray micro-tomography for structural analysis, as well as synchrotron X-ray fluorescence to map the elemental distribution in the infiltrated composite. A hybrid half-cell was constructed and cycled as proof of principle, and it showed good stability. In addition, a bilayer structure consisting of a porous layer combined with a dense LLZO film was successfully made as a prototype of an all solid-state battery. Finally, a mathematical model was established to propose optimized scaffold structures for battery performance.
Friction stir welding (FSW) is a relatively new joining technique particularly for aluminium alloys that are difficult to fusion weld. In this study, the geometry of the weld has been investigated and modelled using Pearson's functions. It has been demonstrated that the Pearson's parameters (mean, standard deviation, skewness, kurtosis and geometric constant) can be used to characterize the weld geometry and the tensile strength of the weld assembly. Pearson's parameters and process parameters are strongly correlated allowing to define a control process procedure for FSW assemblies which make radiographic or ultrasonic controls unnecessary. Finally, an optimisation using a Generalized Gradient Method allows to determine the geometry of the weld which maximises the assembly tensile strength.
In this research, an automatic blade forging die shape optimisation system was developed by using direct compensation and flash trimming algorithms and integrating with the DEFORM 3D software package. To validate the developed system, a 3D blade forging case problem was simulated and optimised with and without the consideration of trimming simulation. The results were compared with actual measurement data of the forged aerofoil blade with excellent results obtained with the fast trimming simulation procedure used.
A two-dimensional model is used to study the geometrical effects of a nonmagnetic (NM) nanowire upon a spin-polarised electron current in a lateral spin-valve structure. We found that the implemented ratchet shapes at the centre of the NM have a crucial effect on the diffusive rate for up- and down-spin electrons along the wire, which leads to the amplification of non-local spin-current signals. By using our simple model, the geometries have been optimised. The calculated spin-current signals are in good qualitative agreement with our recent experimental results [Abdullah et al., J. Phys. D: Appl. Phys. 47, 482001(FTC) (2014)]. Our model may be very useful to evaluate such a geometrical effect on spin-polarised electron transport.
Phenotype heterogeneity in cancer cell populations, be it of genetic, epigenetic or stochastic origin, has been identified as a main source of resistance to drug treatments and a major source of therapeutic failures in cancers. The molecular mechanisms of drug resistance are partly understood at the single cell level (e.g., overexpression of ABC transporters or of detoxication enzymes), but poorly predictable in tumours, where they are hypothesised to rely on heterogeneity at the cell population scale, which is thus the right level to describe cancer growth and optimise its control by therapeutic strategies in the clinic. We review a few results from the biological literature on the subject, and from mathematical models that have been published to predict and control evolution towards drug resistance in cancer cell populations. We propose, based on the latter, optimisation strategies of combined treatments to limit emergence of drug resistance to cytotoxic drugs in cancer cell populations, in the monoclonal situation, which limited as it is still retains consistent features of cell population heterogeneity. The polyclonal situation, that may be understood as “bet hedging” of the tumour, thus protecting itself from different sources of drug insults, may lie beyond such strategies and will need further developments. In the monoclonal situation, we have designed an optimised therapeutic strategy relying on a scheduled combination of cytotoxic and cytostatic treatments that can be adapted to different situations of cancer treatments. Finally, we review arguments for biological theoretical frameworks proposed at different time and development scales, the so-called atavistic model (diachronic view relying on Darwinian genotype selection in the coursof billions of years) and the Waddington-like epigenetic landscape endowed with evolutionary quasi-potential (synchronic view relying on Lamarckian phenotype instruction of a given genome by reversible mechanisms), to represent evolution towards heterogeneity, possibly polyclonal, in cancer cell populations and propose innovative directions for therapeutic strategies based on such frameworks.
Neural Ordinary Differential Equations (ODEs) are a promising approach to learn dynamical models from time-series data in science and engineering applications. This work aims at learning neural ODEs for stiff systems, which are usually raised from chemical kinetic modeling in chemical and biological systems. We first show the challenges of learning neural ODEs in the classical stiff ODE systems of Robertson’s problem and propose techniques to mitigate the challenges associated with scale separations in stiff systems. We then present successful demonstrations in stiff systems of Robertson’s problem and an air pollution problem. The demonstrations show that the usage of deep networks with rectified activations, proper scaling of the network outputs as well as loss functions, and stabilized gradient calculations are the key techniques enabling the learning of stiff neural ODEs. The success of learning stiff neural ODEs opens up possibilities of using neural ODEs in applications with widely varying time-scales, such as chemical dynamics in energy conversion, environmental engineering, and life sciences.
The residential sector accounts for 25% of global primary energy consumption. Two methods have previously been proposed to reduce residential energy use associated with the provision of occupant thermal comfort: 1. Occupancy-based HVAC control, operating systems only during confirmed occupancy, and 2. model predictive control (MPC), harnessing a mathematical model and forecasts to find optimal operating strategies. Previous studies estimate the average energy savings of the two methods individually in the range of 21% and 16%, respectively. The research presented herein was carried out to evaluate the energy savings potential in residential buildings by combining both approaches across different climates, house vintages, and occupancy patterns. Occupancy and eight different physical modalities (e.g. CO2 and VOC) data were collected from five homes for time periods of 4–9 weeks. Collected data sets were used to train occupancy prediction models suggested by an extensive literature survey of occupancy model types. The trained prediction models were combined with MPC and detailed EnergyPlus building simulation models to evaluate residential building performance in terms of annual energy savings and thermal comfort, along with discomfort exceedance metrics. Multiple home types and regions were analyzed to understand regional and climate-dependent potential. Based on actual field data, the occupancy models had a prediction inaccuracy between 8% and 35% across the investigated homes. Average occupancy for the collected data ranged from 56% to 86%, a typical range reported in the literature. Building simulations were conducted for three control scenarios: conventional thermostatic control, occupancy-based, and occupancy-based MPC. The results indicate that all advanced strategies improve upon the conventional control, with some scenarios cutting energy use in half with only occasional incurrence of discomfort. The findings indicate that occupancy-aware model predictive residential building control has the potential to drastically reduce energy use and associated emissions while maintaining occupant comfort for both new and existing buildings.
Using the trace minimization algorithm, we carried out an exact calculation of entanglement in a 19-site two-dimensional transverse Ising model. This model consists of a set of localized spin-(1/2) particles in a two-dimensional triangular lattice coupled through exchange interaction J and subject to an external magnetic field of strength h. We demonstrate, for such a class of two-dimensional magnetic systems, that entanglement can be controlled and tuned by varying the parameter {lambda}=h/J in the Hamiltonian and by introducing impurities into the systems. Examining the derivative of the concurrence as a function of {lambda} shows that the system exhibits a quantum phase transition at about {lambda}{sub c}=3.01, a transition induced by quantum fluctuations at the absolute zero of temperature.
We study a class of nonsupersymmetric SO(10) grand-unified scenarios where the first stage of the symmetry breaking is driven by the vacuum expectation values of the 45-dimensional adjoint representation. Three-decade-old results claim that such a Higgs setting may lead exclusively to the flipped SU(5) x U(1) intermediate stage. We show that this conclusion is actually an artifact of the tree-level potential. The study of the accidental global symmetries emerging in various limits of the scalar potential offers a simple understanding of the tree-level result and a rationale for the drastic impact of quantum corrections. We scrutinize in detail the simplest and paradigmatic case of the 45{sub H}+16{sub H} Higgs sector triggering the breaking of SO(10) to the standard electroweak model. We show that the minimization of the one-loop effective potential allows for intermediate SU(4){sub C} x SU(2){sub L} x U(1){sub R} and SU(3){sub c} x SU(2){sub L} x SU(2){sub R} x U(1){sub B-L} symmetric stages as well. These are the options favored by gauge unification. Our results, that apply whenever the SO(10) breaking is triggered by <45{sub H}>, open the path for hunting the simplest realistic scenario of nonsupersymmetric SO(10) grand unification.