Reinforcement learning in dendritic structures

The discovery of binary dendritic events such as local NMDA spikes in dendritic subbranches led to the suggestion that dendritic trees could be computationally equivalent to a 2-layer network of point neurons, with a single output unit represented by the soma, and input units represented by the dendritic branches. Although this interpretation endows a neuron with a high computational power, it is functionally not clear why nature would have preferred the dendritic solution with a single but complex neuron, as opposed to the network solution with many but simple units. We show that the dendritic solution has a distinguished advantage over the network solution when considering different learning tasks. Its key property is that the dendritic branches receive an immediate feedback from the somatic output spike, while in the corresponding network architecture the feedback would require additional backpropagating connections to the input units. Assuming a reinforcement learning scenario we formally derive a learning rule for the synaptic contacts on the individual dendritic trees which depends on the presynaptic activity, the local NMDA spikes, the somatic action potential, and a delayed reinforcement signal. We test the model for two scenarios: the learning of binary classifications and of precise spike timings. We show that the immediate feedback represented by the backpropagating action potential supplies the individual dendritic branches with enough information to efficiently adapt their synapses and to speed up the learning process.

Reinforcement learning in dendritic structures

The discovery of binary dendritic events such as local NMDA spikes in dendritic subbranches led to the suggestion that dendritic trees could be computationally equivalent to a 2-layer network of point neurons, with a single output unit represented by the soma, and input units represented by the dendritic branches. Although this interpretation endows a neuron with a high computational power, it is functionally not clear why nature would have preferred the dendritic solution with a single but complex neuron, as opposed to the network solution with many but simple units. We show that the dendritic solution has a distinguished advantage over the network solution when considering different learning tasks. Its key property is that the dendritic branches receive an immediate feedback from the somatic output spike, while in the corresponding network architecture the feedback would require additional backpropagating connections to the input units. Assuming a reinforcement learning scenario we formally derive a learning rule for the synaptic contacts on the individual dendritic trees which depends on the presynaptic activity, the local NMDA spikes, the somatic action potential, and a delayed reinforcement signal. We test the model for two scenarios: the learning of binary classifications and of precise spike timings. We show that the immediate feedback represented by the backpropagating action potential supplies the individual dendritic branches with enough information to efficiently adapt their synapses and to speed up the learning process.

Can wild ungulate carcasses provide enough biomass to maintain avian scavenger populations? An empirical assessment using a bio-inspired computational model

Background The reduction in the amount of food available for European avian scavengers as a consequence of restrictive public health policies is a concern for managers and conservationists. Since 2002, the application of several sanitary regulations has limited the availability of feeding resources provided by domestic carcasses, but theoretical studies assessing whether the availability of food resources provided by wild ungulates are enough to cover energetic requirements are lacking. Methodology/Findings We assessed food provided by a wild ungulate population in two areas of NE Spain inhabited by three vulture species and developed a P System computational model to assess the effects of the carrion resources provided on their population dynamics. We compared the real population trend with to a hypothetical scenario in which only food provided by wild ungulates was available. Simulation testing of the model suggests that wild ungulates constitute an important food resource in the Pyrenees and the vulture population inhabiting this area could grow if only the food provided by wild ungulates would be available. On the contrary, in the Pre-Pyrenees there is insufficient food to cover the energy requirements of avian scavenger guilds, declining sharply if biomass from domestic animals would not be available. Conclusions/Significance Our results suggest that public health legislation can modify scavenger population trends if a large number of domestic ungulate carcasses disappear from the mountains. In this case, food provided by wild ungulates could be not enough and supplementary feeding could be necessary if other alternative food resources are not available (i.e. the reintroduction of wild ungulates), preferably in European Mediterranean scenarios sharing similar and socio-economic conditions where there are low densities of wild ungulates. Managers should anticipate the conservation actions required by assessing food availability and the possible scenarios in order to make the most suitable decisions.

A Computational Model for Processing Lexical Information"

The group analysed some syntactic and phonological phenomena that presuppose the existence of interrelated components within the lexicon, which motivate the assumption that there are some sublexicons within the global lexicon of a speaker. This result is confirmed by experimental findings in neurolinguistics. Hungarian speaking agrammatic aphasics were tested in several ways, the results showing that the sublexicon of closed-class lexical items provides a highly automated complex device for processing surface sentence structure. Analysing Hungarian ellipsis data from a semantic-syntactic aspect, the group established that the lexicon is best conceived of being as split into at least two main sublexicons: the store of semantic-syntactic feature bundles and a separate store of sound forms. On this basis they proposed a format for representing open-class lexical items whose meanings are connected via certain semantic relations. They also proposed a new classification of verbs to account for the contribution of the aspectual reading of the sentence depending on the referential type of the argument, and a new account of the syntactic and semantic behaviour of aspectual prefixes. The partitioned sets of lexical items are sublexicons on phonological grounds. These sublexicons differ in terms of phonotactic grammaticality. The degrees of phonotactic grammaticality are tied up with the problem of psychological reality, of how many degrees of this native speakers are sensitive to. The group developed a hierarchical construction network as an extension of the original General Inheritance Network formalism and this framework was then used as a platform for the implementation of the grammar fragments.

Computational Techniques for Spatial Logistic Regression with Large Datasets

In epidemiological work, outcomes are frequently non-normal, sample sizes may be large, and effects are often small. To relate health outcomes to geographic risk factors, fast and powerful methods for fitting spatial models, particularly for non-normal data, are required. We focus on binary outcomes, with the risk surface a smooth function of space. We compare penalized likelihood models, including the penalized quasi-likelihood (PQL) approach, and Bayesian models based on fit, speed, and ease of implementation. A Bayesian model using a spectral basis representation of the spatial surface provides the best tradeoff of sensitivity and specificity in simulations, detecting real spatial features while limiting overfitting and being more efficient computationally than other Bayesian approaches. One of the contributions of this work is further development of this underused representation. The spectral basis model outperforms the penalized likelihood methods, which are prone to overfitting, but is slower to fit and not as easily implemented. Conclusions based on a real dataset of cancer cases in Taiwan are similar albeit less conclusive with respect to comparing the approaches. The success of the spectral basis with binary data and similar results with count data suggest that it may be generally useful in spatial models and more complicated hierarchical models.

Filtration and oxidation characteristics of a diesel oxidation catalyst and a catalyzed particulate filter : development of a 1-D 2-layer model

The emissions, filtration and oxidation characteristics of a diesel oxidation catalyst (DOC) and a catalyzed particulate filter (CPF) in a Johnson Matthey catalyzed continuously regenerating trap (CCRT ®) were studied by using computational models. Experimental data needed to calibrate the models were obtained by characterization experiments with raw exhaust sampling from a Cummins ISM 2002 engine with variable geometry turbocharging (VGT) and programmed exhaust gas recirculation (EGR). The experiments were performed at 20, 40, 60 and 75% of full load (1120 Nm) at rated speed (2100 rpm), with and without the DOC upstream of the CPF. This was done to study the effect of temperature and CPF-inlet NO2 concentrations on particulate matter oxidation in the CCRT ®. A previously developed computational model was used to determine the kinetic parameters describing the oxidation characteristics of HCs, CO and NO in the DOC and the pressure drop across it. The model was calibrated at five temperatures in the range of 280 – 465° C, and exhaust volumetric flow rates of 0.447 – 0.843 act-m3/sec. The downstream HCs, CO and NO concentrations were predicted by the DOC model to within ±3 ppm. The HCs and CO oxidation kinetics in the temperature range of 280 - 465°C and an exhaust volumetric flow rate of 0.447 - 0.843 act-m3/sec can be represented by one ’apparent’ activation energy and pre-exponential factor. The NO oxidation kinetics in the same temperature and exhaust flow rate range can be represented by ’apparent’ activation energies and pre-exponential factors in two regimes. The DOC pressure drop was always predicted within 0.5 kPa by the model. The MTU 1-D 2-layer CPF model was enhanced in several ways to better model the performance of the CCRT ®. A model to simulate the oxidation of particulate inside the filter wall was developed. A particulate cake layer filtration model which describes particle filtration in terms of more fundamental parameters was developed and coupled to the wall oxidation model. To better model the particulate oxidation kinetics, a model to take into account the NO2 produced in the washcoat of the CPF was developed. The overall 1-D 2-layer model can be used to predict the pressure drop of the exhaust gas across the filter, the evolution of particulate mass inside the filter, the particulate mass oxidized, the filtration efficiency and the particle number distribution downstream of the CPF. The model was used to better understand the internal performance of the CCRT®, by determining the components of the total pressure drop across the filter, by classifying the total particulate matter in layer I, layer II, the filter wall, and by the means of oxidation i.e. by O2, NO2 entering the filter and by NO2 being produced in the filter. The CPF model was calibrated at four temperatures in the range of 280 – 465 °C, and exhaust volumetric flow rates of 0.447 – 0.843 act-m3/sec, in CPF-only and CCRT ® (DOC+CPF) configurations. The clean filter wall permeability was determined to be 2.00E-13 m2, which is in agreement with values in the literature for cordierite filters. The particulate packing density in the filter wall had values between 2.92 kg/m3 - 3.95 kg/m3 for all the loads. The mean pore size of the catalyst loaded filter wall was found to be 11.0 µm. The particulate cake packing densities and permeabilities, ranged from 131 kg/m3 - 134 kg/m3, and 0.42E-14 m2 and 2.00E-14 m2 respectively, and are in agreement with the Peclet number correlations in the literature. Particulate cake layer porosities determined from the particulate cake layer filtration model ranged between 0.841 and 0.814 and decreased with load, which is about 0.1 lower than experimental and more complex discrete particle simulations in the literature. The thickness of layer I was kept constant at 20 µm. The model kinetics in the CPF-only and CCRT ® configurations, showed that no ’catalyst effect’ with O2 was present. The kinetic parameters for the NO2-assisted oxidation of particulate in the CPF were determined from the simulation of transient temperature programmed oxidation data in the literature. It was determined that the thermal and NO2 kinetic parameters do not change with temperature, exhaust flow rate or NO2 concentrations. However, different kinetic parameters are used for particulate oxidation in the wall and on the wall. Model results showed that oxidation of particulate in the pores of the filter wall can cause disproportionate decreases in the filter pressure drop with respect to particulate mass. The wall oxidation model along with the particulate cake filtration model were developed to model the sudden and rapid decreases in pressure drop across the CPF. The particulate cake and wall filtration models result in higher particulate filtration efficiencies than with just the wall filtration model, with overall filtration efficiencies of 98-99% being predicted by the model. The pre-exponential factors for oxidation by NO2 did not change with temperature or NO2 concentrations because of the NO2 wall production model. In both CPF-only and CCRT ® configurations, the model showed NO2 and layer I to be the dominant means and dominant physical location of particulate oxidation respectively. However, at temperatures of 280 °C, NO2 is not a significant oxidizer of particulate matter, which is in agreement with studies in the literature. The model showed that 8.6 and 81.6% of the CPF-inlet particulate matter was oxidized after 5 hours at 20 and 75% load in CCRT® configuration. In CPF-only configuration at the same loads, the model showed that after 5 hours, 4.4 and 64.8% of the inlet particulate matter was oxidized. The increase in NO2 concentrations across the DOC contributes significantly to the oxidation of particulate in the CPF and is supplemented by the oxidation of NO to NO2 by the catalyst in the CPF, which increases the particulate oxidation rates. From the model, it was determined that the catalyst in the CPF modeslty increases the particulate oxidation rates in the range of 4.5 – 8.3% in the CCRT® configuration. Hence, the catalyst loading in the CPF of the CCRT® could possibly be reduced without significantly decreasing particulate oxidation rates leading to catalyst cost savings and better engine performance due to lower exhaust backpressures.

Discrete element methods for asphalt concrete : development and application of user-defined microstructural models and a viscoelastic micromechanical model

As an important Civil Engineering material, asphalt concrete (AC) is commonly used to build road surfaces, airports, and parking lots. With traditional laboratory tests and theoretical equations, it is a challenge to fully understand such a random composite material. Based on the discrete element method (DEM), this research seeks to develop and implement computer models as research approaches for improving understandings of AC microstructure-based mechanics. In this research, three categories of approaches were developed or employed to simulate microstructures of AC materials, namely the randomly-generated models, the idealized models, and image-based models. The image-based models were recommended for accurately predicting AC performance, while the other models were recommended as research tools to obtain deep insight into the AC microstructure-based mechanics. A viscoelastic micromechanical model was developed to capture viscoelastic interactions within the AC microstructure. Four types of constitutive models were built to address the four categories of interactions within an AC specimen. Each of the constitutive models consists of three parts which represent three different interaction behaviors: a stiffness model (force-displace relation), a bonding model (shear and tensile strengths), and a slip model (frictional property). Three techniques were developed to reduce the computational time for AC viscoelastic simulations. It was found that the computational time was significantly reduced to days or hours from years or months for typical three-dimensional models. Dynamic modulus and creep stiffness tests were simulated and methodologies were developed to determine the viscoelastic parameters. It was found that the DE models could successfully predict dynamic modulus, phase angles, and creep stiffness in a wide range of frequencies, temperatures, and time spans. Mineral aggregate morphology characteristics (sphericity, orientation, and angularity) were studied to investigate their impacts on AC creep stiffness. It was found that aggregate characteristics significantly impact creep stiffness. Pavement responses and pavement-vehicle interactions were investigated by simulating pavement sections under a rolling wheel. It was found that wheel acceleration, steadily moving, and deceleration significantly impact contact forces. Additionally, summary and recommendations were provided in the last chapter and part of computer programming codes wree provided in the appendixes.

Integration of computational models and experimental characterization to study internal frost damage in cementitious materials

The objective of this doctoral research is to investigate the internal frost damage due to crystallization pore pressure in porous cement-based materials by developing computational and experimental characterization tools. As an essential component of the U.S. infrastructure system, the durability of concrete has significant impact on maintenance costs. In cold climates, freeze-thaw damage is a major issue affecting the durability of concrete. The deleterious effects of the freeze-thaw cycle depend on the microscale characteristics of concrete such as the pore sizes and the pore distribution, as well as the environmental conditions. Recent theories attribute internal frost damage of concrete is caused by crystallization pore pressure in the cold environment. The pore structures have significant impact on freeze-thaw durability of cement/concrete samples. The scanning electron microscope (SEM) and transmission X-ray microscopy (TXM) techniques were applied to characterize freeze-thaw damage within pore structure. In the microscale pore system, the crystallization pressures at sub-cooling temperatures were calculated using interface energy balance with thermodynamic analysis. The multi-phase Extended Finite Element Modeling (XFEM) and bilinear Cohesive Zone Modeling (CZM) were developed to simulate the internal frost damage of heterogeneous cement-based material samples. The fracture simulation with these two techniques were validated by comparing the predicted fracture behavior with the captured damage from compact tension (CT) and single-edge notched beam (SEB) bending tests. The study applied the developed computational tools to simulate the internal frost damage caused by ice crystallization with the two dimensional (2-D) SEM and three dimensional (3-D) reconstructed SEM and TXM digital samples. The pore pressure calculated from thermodynamic analysis was input for model simulation. The 2-D and 3-D bilinear CZM predicted the crack initiation and propagation within cement paste microstructure. The favorably predicted crack paths in concrete/cement samples indicate the developed bilinear CZM techniques have the ability to capture crack nucleation and propagation in cement-based material samples with multiphase and associated interface. By comparing the computational prediction with the actual damaged samples, it also indicates that the ice crystallization pressure is the main mechanism for the internal frost damage in cementitious materials.

Development of a continuum mechanics model of passive skeletal muscle

Skeletal muscle force evaluation is difficult to implement in a clinical setting. Muscle force is typically assessed through either manual muscle testing, isokinetic/isometric dynamometry, or electromyography (EMG). Manual muscle testing is a subjective evaluation of a patient’s ability to move voluntarily against gravity and to resist force applied by an examiner. Muscle testing using dynamometers adds accuracy by quantifying functional mechanical output of a limb. However, like manual muscle testing, dynamometry only provides estimates of the joint moment. EMG quantifies neuromuscular activation signals of individual muscles, and is used to infer muscle function. Despite the abundance of work performed to determine the degree to which EMG signals and muscle forces are related, the basic problem remains that EMG cannot provide a quantitative measurement of muscle force. Intramuscular pressure (IMP), the pressure applied by muscle fibers on interstitial fluid, has been considered as a correlate for muscle force. Numerous studies have shown that an approximately linear relationship exists between IMP and muscle force. A microsensor has recently been developed that is accurate, biocompatible, and appropriately sized for clinical use. While muscle force and pressure have been shown to be correlates, IMP has been shown to be non-uniform within the muscle. As it would not be practicable to experimentally evaluate how IMP is distributed, computational modeling may provide the means to fully evaluate IMP generation in muscles of various shapes and operating conditions. The work presented in this dissertation focuses on the development and validation of computational models of passive skeletal muscle and the evaluation of their performance for prediction of IMP. A transversly isotropic, hyperelastic, and nearly incompressible model will be evaluated along with a poroelastic model.

Computational study of internal and external condensing flows and experimental synthesis to investigate their attainability and stability in ground-based and space-based environments

This doctoral thesis presents the computational work and synthesis with experiments for internal (tube and channel geometries) as well as external (flow of a pure vapor over a horizontal plate) condensing flows. The computational work obtains accurate numerical simulations of the full two dimensional governing equations for steady and unsteady condensing flows in gravity/0g environments. This doctoral work investigates flow features, flow regimes, attainability issues, stability issues, and responses to boundary fluctuations for condensing flows in different flow situations. This research finds new features of unsteady solutions of condensing flows; reveals interesting differences in gravity and shear driven situations; and discovers novel boundary condition sensitivities of shear driven internal condensing flows. Synthesis of computational and experimental results presented here for gravity driven in-tube flows lays framework for the future two-phase component analysis in any thermal system. It is shown for both gravity and shear driven internal condensing flows that steady governing equations have unique solutions for given inlet pressure, given inlet vapor mass flow rate, and fixed cooling method for condensing surface. But unsteady equations of shear driven internal condensing flows can yield different “quasi-steady” solutions based on different specifications of exit pressure (equivalently exit mass flow rate) concurrent to the inlet pressure specification. This thesis presents a novel categorization of internal condensing flows based on their sensitivity to concurrently applied boundary (inlet and exit) conditions. The computational investigations of an external shear driven flow of vapor condensing over a horizontal plate show limits of applicability of the analytical solution. Simulations for this external condensing flow discuss its stability issues and throw light on flow regime transitions because of ever-present bottom wall vibrations. It is identified that laminar to turbulent transition for these flows can get affected by ever present bottom wall vibrations. Detailed investigations of dynamic stability analysis of this shear driven external condensing flow result in the introduction of a new variable, which characterizes the ratio of strength of the underlying stabilizing attractor to that of destabilizing vibrations. Besides development of CFD tools and computational algorithms, direct application of research done for this thesis is in effective prediction and design of two-phase components in thermal systems used in different applications. Some of the important internal condensing flow results about sensitivities to boundary fluctuations are also expected to be applicable to flow boiling phenomenon. Novel flow sensitivities discovered through this research, if employed effectively after system level analysis, will result in the development of better control strategies in ground and space based two-phase thermal systems.

Computational fluids domain reduction to a simplified fluid network

The primary goal of this project is to demonstrate the practical use of data mining algorithms to cluster a solved steady-state computational fluids simulation (CFD) flow domain into a simplified lumped-parameter network. A commercial-quality code, “cfdMine” was created using a volume-weighted k-means clustering that that can accomplish the clustering of a 20 million cell CFD domain on a single CPU in several hours or less. Additionally agglomeration and k-means Mahalanobis were added as optional post-processing steps to further enhance the separation of the clusters. The resultant nodal network is considered a reduced-order model and can be solved transiently at a very minimal computational cost. The reduced order network is then instantiated in the commercial thermal solver MuSES to perform transient conjugate heat transfer using convection predicted using a lumped network (based on steady-state CFD). When inserting the lumped nodal network into a MuSES model, the potential for developing a “localized heat transfer coefficient” is shown to be an improvement over existing techniques. Also, it was found that the use of the clustering created a new flow visualization technique. Finally, fixing clusters near equipment newly demonstrates a capability to track temperatures near specific objects (such as equipment in vehicles).

COMPUTATIONAL ANALYSIS OF THE AEROELASTIC DYNAMICS OF A WIND TURBINE BLADE WITH VARIABLE-SPEED STALL CONTROL

The accuracy of simulating the aerodynamics and structural properties of the blades is crucial in the wind-turbine technology. Hence the models used to implement these features need to be very precise and their level of detailing needs to be high. With the variety of blade designs being developed the models should be versatile enough to adapt to the changes required by every design. We are going to implement a combination of numerical models which are associated with the structural and the aerodynamic part of the simulation using the computational power of a parallel HPC cluster. The structural part models the heterogeneous internal structure of the beam based on a novel implementation of the Generalized Timoshenko Beam Model Technique.. Using this technique the 3-D structure of the blade is reduced into a 1-D beam which is asymptotically equivalent. This reduces the computational cost of the model without compromising its accuracy. This structural model interacts with the Flow model which is a modified version of the Blade Element Momentum Theory. The modified version of the BEM accounts for the large deflections of the blade and also considers the pre-defined structure of the blade. The coning, sweeping of the blade, tilt of the nacelle and the twist of the sections along the blade length are all computed by the model which aren’t considered in the classical BEM theory. Each of these two models provides feedback to the other and the interactive computations lead to more accurate outputs. We successfully implemented the computational models to analyze and simulate the structural and aerodynamic aspects of the blades. The interactive nature of these models and their ability to recompute data using the feedback from each other makes this code more efficient than the commercial codes available. In this thesis we start off with the verification of these models by testing it on the well-known benchmark blade for the NREL-5MW Reference Wind Turbine, an alternative fixed-speed stall-controlled blade design proposed by Delft University, and a novel alternative design that we proposed for a variable-speed stall-controlled turbine, which offers the potential for more uniform power control and improved annual energy production.. To optimize the power output of the stall-controlled blade we modify the existing designs and study their behavior using the aforementioned aero elastic model.

COMPUTATIONAL STUDY OF MICROSTRUCTURE-PROPERTYMECHANISM RELATIONS IN FERROIC COMPOSITES

Ferroic materials, as notable members of smart materials, have been widely used in applications that perform sensing, actuation and control. The macroscopic property change of ferroic materials may become remarkably large during ferroic phase transition, leading to the fact that the macroscopic properties can be tuned by carefully applying a suitable external field (electric, magnetic, stress). To obtain an enhancement in physical and/or mechanical properties, different kinds of ferroic composites have been fabricated. The properties of a ferroic composite are determined not only by the properties and relative amounts of the constituent phases, but also by the microstructure of individual phase such as the phase connectivity, phase size, shape and spatial arrangement. This dissertation mainly focuses on the computational study of microstructure – property – mechanism relations in two representative ferroic composites, i.e., two-phase particulate magnetoelectric (ME) composite and polymer matrix ferroelectric composite. The former is a great example of ferroic composite exhibiting a new property and functionality that neither of the constituent phases possesses individually. The latter well represents the kind of ferroic composites having property combinations that are better than the existing materials. Phase field modeling was employed as the computing tool, and the required models for ferroic composites were developed based on existing models for monolithic materials. Extensive computational simulations were performed to investigate the microstructure-property relations and the underlying mechanism in ferroic composites. In particulate, it is found that for ME composite 0-3 connectivity (isolated magnetostrictive phase) is necessary to exhibit ME effect, and small but finite electrical conductivity of isolated magnetic phase can beneficially enhance ME effect. It is revealed that longitudinal and transverse ME coefficients of isotropic 0-3 particulate composites can be effectively tailored by controlling magnetic domain structures without resort to anisotropic two-phase microstructures. Simulations also show that the macroscopic properties of the ferroelectricpolymer composites critically depend on the ferroelectric phase connectivity while are not sensitive to the sizes and internal grain structures of the ceramic particles. Texturing is found critical to exploit the paraelectric«ferroelectric phase transition and nonlinear polarization behavior in paraelectric polycrystal and its polymer matrix composite. Additionally, a Diffuse Interface Field model was developed to simulate packing and motion in liquid phase which is promising for studying the fabrication of particulatepolymer composites.

XML for Model Specification in Neuroscience

One of the main roles of the Neural Open Markup Language, NeuroML, is to facilitate cooperation in building, simulating, testing and publishing models of channels, neurons and networks of neurons. MorphML, which was developed as a common format for exchange of neural morphology data, is distributed as part of NeuroML but can be used as a stand-alone application. In this collection of tutorials and workshop summary, we provide an overview of these XML schemas and provide examples of their use in down-stream applications. We also summarize plans for the further development of XML specifications for modeling channels, channel distributions, and network connectivity.