Maintain Model Accuracy Research Articles

1xN Pattern for Pruning Convolutional Neural Networks.

Though network pruning receives popularity in reducing the complexity of convolutional neural networks (CNNs), it remains an open issue to concurrently maintain model accuracy as well as achieve significant speedups on general CPUs. In this paper, we propose a novel 1×N pruning pattern to break this limitation. In particular, consecutive N output kernels with the same input channel index are grouped into one block, which serves as a basic pruning granularity of our pruning pattern. Our 1×N pattern prunes these blocks considered unimportant. We also provide a workflow of filter rearrangement that first rearranges the weight matrix in the output channel dimension to derive more influential blocks for accuracy improvements and then applies similar rearrangement to the next-layer weights in the input channel dimension to ensure correct convolutional operations. Moreover, the output computation after our 1×N pruning can be realized via a parallelized block-wise vectorized operation, leading to significant speedups on general CPUs. The efficacy of our pruning pattern is proved with experiments on ILSVRC-2012. For example, given the pruning rate of 50% and N=4, our pattern obtains about 3.0% improvements over filter pruning in the top-1 accuracy of MobileNet-V2. Meanwhile, it obtains 56.04ms inference savings on Cortex-A7 CPU over weight pruning. Our project is made available at https://github.com/lmbxmu/1xN.

Abstract 242: Transferring diagnostic and prognostic molecular models across technological platforms

Abstract The reproducibility of results obtained using RNA data across labs is a major hurdle in cancer research. Difference in library preparation methods and gene expression quantification platforms prevent the application of trained models to new data across labs. SpinAdapt is a novel unsupervised domain adaptation algorithm that enables the transfer of existing molecular models across labs and technological platforms, without requiring re-training or calibration of existing models for future prospective data. Furthermore, SpinAdapt uses summary statistics (independent latent space representations) to calculate data corrections, rather than requiring full data access. This allows for transfer of molecular models across sequencing platforms and between labs without loss of data ownership or compromise of data privacy. To evaluate SpinAdapt, we performed two sets of experiments: A) We transferred molecular tumor subtype classifiers across four pairs of publicly available cancer datasets (bladder, breast, colorectal, pancreatic), covering 4,076 samples across 18 different tumor subtypes and three technological platforms (RNASeq, Affymetrix U133plus2, and HE1ST). For each pair of datasets we trained a subtype classifier on one dataset (target) according to well-accepted subtyping annotations (Zea Tan et al. 2019; Prat et al. 2012; Guinney et al. 2015; Bailey et al. 2016), and then evaluated the classifier accuracy on the other dataset (source). For each tumor subtype, we quantified the classification performance using mean AUC score across random subsets of the source dataset, where each subset was corrected using SpinAdapt. We aggregated performance across all subtypes and report the average mean AUC score for each cancer type: bladder 0.95, breast 0.98, colorectal 0.98, pancreatic 0.96; demonstrating high accuracy on all diagnostic tasks. B) To demonstrate the transferability of prognostic models, we trained five Cox survival models on five target cancer datasets respectively (breast, lung, colorectal, liver, pancreatic) ranging from 186 to 2,919 RNASeq samples. We used SpinAdapt to adapt five source cancer datasets to the target datasets, ranging from 226 to 1,038 samples across different platforms (RNASeq, Affymetrix U133Plus2 and HG-U133A, Illumina HumanHT-12v4). For every cancer type, we trained a Cox model on the target dataset, and measured its performance by predicting survival risk on the corresponding adapted source dataset. We show high survival prediction accuracy for all datasets (Log-rank P-values and c-index): lung [1e-6, 01.661], breast [5e-5, 0.626], liver [1e-4, 0.708], pancreatic [2e-4, 0.629], colorectal [9e-4, 0.661]. SpinAdapt transferred diagnostic and prognostic models over 14 cancer datasets covering 7,146 samples across six different cancer types and various platforms (RNASeq, microarray), while maintaining model accuracy and statistical significance. Citation Format: Talal Ahmed, Stephane Wenric, Mark Carty, Rafael Pelossof. Transferring diagnostic and prognostic molecular models across technological platforms [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 242.

Prediction of regional wildfire activity in the probabilistic Bayesian framework of Firelihood.

Modeling wildfire activity is crucial for informing science-based risk management and understanding the spatiotemporal dynamics of fire-prone ecosystems worldwide. Models help disentangle the relative influences of different factors, understand wildfire predictability, and provide insights into specific events. Here, we develop Firelihood, a two-component, Bayesian, hierarchically structured, probabilistic model of daily fire activity, which is modeled as the outcome of a marked point process: individual fires are the points (occurrence component), and fire sizes are the marks (size component). The space-time Poisson model for occurrence is adjusted to gridded fire counts using the integrated nested Laplace approximation (INLA) combined with the stochastic partial differential equation (SPDE) approach. The size model is based on piecewise-estimated Pareto and generalized Pareto distributions, adjusted with INLA. The Fire Weather Index (FWI) and forest area are the main explanatory variables. Temporal and spatial residuals are included to improve the consistency of the relationship between weather and fire occurrence. The posterior distribution of the Bayesian model provided 1,000 replications of fire activity that were compared with observations at various temporal and spatial scales in Mediterranean France. The number of fires larger than 1ha across the region was coarsely reproduced at the daily scale, and was more accurately predicted on a weekly basis or longer. The regional weekly total number of larger fires (10-100ha) was predicted as well, but the accuracy degraded with size, as the model uncertainty increased with event rareness. Local predictions of fire numbers or burned areas also required a longer aggregation period to maintain model accuracy. The estimation of fires larger than 1ha was also consistent with observations during the extreme fire season of the 2003 unprecedented heat wave, but the model systematically underrepresented large fires and burned areas, which suggests that the FWI does not consistently rate the actual danger of large fire occurrence during heat waves. Firelihood enabled a novel analysis of the stochasticity underlying fire hazard, and offers a variety of applications, including fire hazard predictions for management and projections in the context of climate change.

On Multi-Event Co-Calibration of Dynamic Model Parameters Using Soft Actor-Critic

Maintaining good quality of transient stability models for power system planning and operational analysis is of great importance. Identification and calibration of bad parameters using PMU measurements that work well for multiple events remains a challenging problem. In this letter, we present a novel parameter calibration method based on off-policy deep reinforcement learning (DRL) algorithm with maximum entropy, soft actor critic (SAC), to automatically tune incorrect parameter sets considering multiple events simultaneously, which can save tremendous labor efforts for maintaining model accuracy and complying with industry standards. The effectiveness of the proposed approach is verified through numerical experiments conducted on a realistic power plant model.

Hierarchical Fuzzy Neural Networks With Privacy Preservation for Heterogeneous Big Data

Heterogeneous big data poses many challenges in machine learning. Its enormous scale, high dimensionality, and inherent uncertainty make almost every aspect of machine learning difficult, from providing enough processing power to maintaining model accuracy to protecting privacy. However, perhaps the most imposing problem is that big data is often interspersed with sensitive personal data. Hence, we propose a privacy-preserving hierarchical fuzzy neural network (PP-HFNN) to address these technical challenges while also alleviating privacy concerns. The network is trained with a two-stage optimization algorithm, and the parameters at low levels of the hierarchy are learned with a scheme based on the well-known alternating direction method of multipliers, which does not reveal local data to other agents. Coordination at high levels of the hierarchy is handled by the alternating optimization method, which converges very quickly. The entire training procedure is scalable, fast and does not suffer from gradient vanishing problems like the methods based on back-propagation. Comprehensive simulations conducted on both regression and classification tasks demonstrate the effectiveness of the proposed model.

An Improved NO Prediction Model for Large Eddy Simulation of Turbulent Combustion

Accurate prediction of nitrogen oxides (NOX) emissions is of great importance in combustion simulations. This work proposes an improved model to predict NO distributions in turbulent flames, based on the large eddy simulation approach and flamelet-progress-variable turbulence combustion model. A separate table for NO reaction rate is constructed to account for both unsteady and nonadiabatic effects on NO production. Different from the existing models, the nonadiabatic effect is considered by specifying different enthalpy defects by a scaling factor in the flamelet calculations. Additionally, it is observed that the transient variation of NO reaction rate with NO concentration exhibits two linear stages. As a result, only three flamelet databases are included to describe the NO reaction rate in a transient process, thereby simplifying the flamelet table while maintaining model accuracy. The proposed NO prediction model is validated in large eddy simulations of turbulent combustion, including the classic Sandia D and the Sydney swirling SM1 flames. The calculated NO variations show good agreement with experimental data, demonstrating the applicability and accuracy of the proposed model.

Model Optimization for the Prediction of Red Wine Phenolic Compounds Using Ultraviolet-Visible Spectra.

The primary objective of this work was to optimize red wine phenolic prediction with models built from wine ultraviolet–visible absorbance spectra. Three major obstacles were addressed to achieve this, namely algorithm selection, spectral multicollinearity, and phenolic evolution over time. For algorithm selection, support vector regression, kernel ridge regression, and kernel partial least squares regression were compared. For multicollinearity, the spectrum of malvidin chloride was used as an external standard for spectral adjustment. For phenolic evolution, spectral data were collected during fermentation as well as once a week for four weeks after fermentation had ended. Support vector regression gave the most accurate predictions among the three algorithms tested. Additionally, malvidin chloride proved a useful standard for phenolic spectral transformation and isolation. As for phenolic evolution, models needed to be calibrated and validated throughout the aging process to ensure predictive accuracy. In short, red wine phenolic prediction by the models built in this work can be realistically achieved, although periodic model re-calibration and expansion from data obtained using known phenolic assays is recommended to maintain model accuracy.

Integrating deep learning models and multiparametric programming

Deep learning models are a class of approximate models that are proven to have strong predictive capabilities for representing complex phenomena. The introduction of deep learning models into an optimization formulation provides a means to reduce the problem complexity and maintain model accuracy. Recently it has been shown that deep learning models in the form of neural networks with rectified linear units can be exactly recast as a mixed-integer linear programming formulation. However, developing the optimal solution of problems involving mixed-integer decisions in online applications remains challenging. Multiparametric programming alleviates the online computational burden of solving an optimization problem involving bounded uncertain parameters. In this work, a strategy is presented to integrate deep learning and multiparametric programming. This integration yields a unified methodology for developing accurate surrogate models based on deep learning and their offline, explicit optimal solution. The proposed strategy is demonstrated on the optimal operation of a chemostat.

Adaptive selective catalytic reduction model development using typical operating data in coal-fired power plants

This study develops an adaptive selective catalytic reduction (SCR) model in a coal-fired power plant with typical operating data to tackle two issues: selecting appropriate samples for model training and maintaining model accuracy under new operating conditions. First, an index of representing the information contained in the operating data of SCR is defined by considering three factors including variation span, distribution status, and information redundancy. Next, the genetic algorithm (GA) is applied to select typical operating data from SCR operational database by maximizing the information index. These data are taken as the training set to develop SCR models and predict NOx emissions with artificial intelligence techniques, including least square support vector machine and artificial neural network. Furthermore, typical operating data are managed adaptively to cover information from new operating conditions, and SCR models are updated according to the data change. SCR models trained with data from other common selections are compared. Results show that the typical operating data selected by GA can contain large information, and the developed models perform better than those trained with data from other selections. In addition, data management and model update can make the model maintain high prediction accuracy when new operating conditions occur.

Application of A Simple Landsat-MODIS Fusion Model to Estimate Evapotranspiration over A Heterogeneous Sparse Vegetation Region

A simple Landsat-MODIS (Moderate Resolution Imaging Spectroradiometer) fusion model was used to generate 30-m resolution evapotranspiration (ET) maps for the 2010 growing season over a heterogeneous sparse vegetation, agricultural region using the METRIC (mapping evapotranspiration with internalized calibration) algorithm. The fusion model performance was evaluated, and experiments were undertaken to investigate the frequency for updating Landsat-MODIS data into the fusion model during the growing season, to maintain model accuracy and reduce computation. Initial evaluation of the fusion model resulted in high bias stemming from the landscape heterogeneity and small landholdings. To reduce the bias, the fusion model was modified to be applicable pixel-wise (i.e., implementing specific pixels for generating outputs), and an NDVI-based (Normalized Difference Vegetation Index) coefficient was added to capture crop phenology. A good agreement that resulted from the comparison of the fused and non-fused maps with root mean square error (RMSE) of 0.15 mm day−1 with coefficient of determination (R2) of 0.83 indicated successful implementation of the modifications. Additionally, the fusion model performance was evaluated against in-situ observation at the pixel level as well as the watershed level to estimate seasonal ET for the growing season. The default METRIC model (Landsat only) yielded relative error (RE) of 31% and RMSE of 2.44 mm day−1, while using the modified fusion model improved the accuracy resulting in RE of 3.5% with RMSE of 0.37 mm day−1. Considering different data frequency update, the optimal fusion experiment (RMSE of 0.61 mm day−1, and RE of 6.5%) required the consideration of the crop phenology and weekly updates in the early growing stage and harvest time, and bi-weekly for the rest of the season. The resulting fusion model for ET output is planned to be a part of ET mapping and irrigation scheduling systems.

Fast and accurate district heating and cooling energy demand and load calculations using reduced-order modelling

Recent developments in building energy models for urban energy simulation are primarily based on bottom-up modelling (N models used for N buildings). This work aims to develop a single assembled model for multiple buildings for convenient use in detailed urban analysis. The proposed model exhibits state-space model formalism, and a state-size reduction technique is applied to maintain model accuracy, even for a low-order representation. To accelerate the calculation time and ensure numerical stability, a direct solver is proposed to eliminate the iterative calculations required in Dymola for annual load calculations. The results of the proposed reduced model are in good agreement with the reference model. For a test case of ten buildings, a 2nd order reduced model (i.e., 2 differential equations) with the proposed direct solver can predict accurately the dynamic energy behaviour, resulting in an error of about 0.43% for the annual loads.

Variable selection using Gaussian process regression-based metrics for high-dimensional model approximation with limited data

In recent years, the importance of computationally efficient surrogate models has been emphasized as the use of high-fidelity simulation models increases. However, high-dimensional models require a lot of samples for surrogate modeling. To reduce the computational burden in the surrogate modeling, we propose an integrated algorithm that incorporates accurate variable selection and surrogate modeling. One of the main strengths of the proposed method is that it requires less number of samples compared with conventional surrogate modeling methods by excluding dispensable variables while maintaining model accuracy. In the proposed method, the importance of selected variables is evaluated using the quality of the model approximated with the selected variables only. Nonparametric probabilistic regression is adopted as the modeling method to deal with inaccuracy caused by using selected variables during modeling. In particular, Gaussian process regression (GPR) is utilized for the modeling because it is suitable for exploiting its model performance indices in the variable selection criterion. Outstanding variables that result in distinctly superior model performance are finally selected as essential variables. The proposed algorithm utilizes a conservative selection criterion and appropriate sequential sampling to prevent incorrect variable selection and sample overuse. Performance of the proposed algorithm is verified with two test problems with challenging properties such as high dimension, nonlinearity, and the existence of interaction terms. A numerical study shows that the proposed algorithm is more effective as the fraction of dispensable variables is high.

Solution of the Nonlinear High-Fidelity Generalized Method of Cells Micromechanics Relations via Order-Reduction Techniques

The High-Fidelity Generalized Method of Cells (HFGMC) is one technique, distinct from traditional finite-element approaches, for accurately simulating nonlinear composite material behavior. In this work, the HFGMC global system of equations for doubly periodic repeating unit cells with nonlinear constituents has been reduced in size through the novel application of a Petrov-Galerkin Proper Orthogonal Decomposition order-reduction scheme in order to improve its computational efficiency. Order-reduced models of an E-glass/Nylon 12 composite led to a 4.8–6.3x speedup in the equation assembly/solution runtime while maintaining model accuracy. This corresponded to a 21–38% reduction in total runtime. The significant difference in assembly/solution and total runtimes was attributed to the evaluation of integration point inelastic field quantities; this step was identical between the unreduced and order-reduced models. Nonetheless, order-reduced techniques offer the potential to significantly improve the computational efficiency of multiscale calculations.

On the current state of flotation modelling for process control

Despite significant effort in modelling and simulating flotation circuits, comprehensive model based control and optimisation implementations on industrial circuits remain scarce. In this paper, the factors preventing more widespread implementation of model-based control and optimisation applications are investigated by focussing on three aspects. Firstly, the critical variables required in a simplified flotation model are identified. Models that are currently used in control, optimisation and supervisory applications are thereafter analysed to determine to what extent the required variables are modelled. Finally, online instrumentation available to support these models are investigated, also including instrumentation that is still under development and not commonly available in commercial applications. Although models used in control applications tend to focus on subsections of the flotation process, there seem to be a good agreement between the required and modelled variables. Model fitting however often relies on extensive sampling campaigns that will need to be repeated regularly to maintain model accuracy. A number of online measurements of sufficient accuracy are still not available to support these models, compromising the long term reliable use of models in online applications. The fact that flotation processes are in many instances not extensively instrumented, constrains online maintenance and adaption of model based solutions further.

An efficient method for fractured shale reservoir history matching: The embedded discrete fracture multi-continuum approach

In this study, we established a more efficient approach for fractured shale reservoir modeling with an emphasis on simplifying and automating the workflow for assisted history matching and uncertainty quantification. The improvement is especially notable for the process of history matching since the fracture geometry and properties can be directly set as parameters to be history matched. The resultant approach not only shows a significant reduction in the computational time while maintaining model accuracy, but also provides an automatic method for modifying the fracture related parameters - a laborious process in the traditional workflow.In the forward reservoir model, we implemented and extended the Embedded Discrete Fracture Model (Embedded DFM) approach for fractures with arbitrary strike and dip angle to a multiple porosity/permeability setting. The fractures are naturally discretized by the boundary of parent matrix grid blocks. Control volumes of fracture segments are generated according to the specific geometry of each of the segments. Three types of non-neighbor connections are then generated, namely the connection between the fracture segment and its parent matrix grid blocks, the connection between two intersecting fracture segments from different fractures, and the connection between two neighbor fracture segments from the same fracture. For each of the non-neighbor connections, transmissibility can be calculated honoring the physics of the flow.In our approach with Embedded Discrete Fracture Multiple-Porosity Model, the matrix is sub-divided into three porosity types, namely organic matrix (kerogen), inorganic matrix and natural fractures, with the necessary physics included for each of the porosity types. The macro fractures are explicitly represented with Embedded DFM. The proposed model provides a coherent method for characterizing the organic matrix, inorganic matrix, micro fractures as well as the hydraulic fractures of shale reservoirs. It offers a computationally efficient approach for modeling the severe heterogeneity due to hydraulic and natural fractures. Compared with traditional discrete fracture models, fewer grid blocks and lower levels of refinement are required. Compared with multiple porosity method, the proposed model has desirable accuracy for the simulation of reservoirs with large scale fractures.In the history matching and uncertainty quantification stage, due to the low efficiency of traditional Markov Chain Monte Carlo (MCMC) method when applied to reservoir history matching, a more advanced algorithm of two-stage MCMC is employed to evaluate the uncertainty for all the parameters. Since no upscaling of the fracture related parameters is required, the reservoir model can be generated by a pre-processor based on the proposed parameter, which maintains the adequacy of a Gaussian distribution assumption. Therefore, the workflow can be completely automated. By incorporating Embedded DFM and multiple porosity/permeability approaches, the improved model facilitates the history matching of fractured shale reservoirs by cutting the total amount of grid blocks, reducing the complexity of the gridding process, as well as improving the accuracy of fluid transportation within and among different porosity types.

StochasticNet in StochasticNet

Deep neural networks have been shown to outperform conventionalstate-of-the-art approaches in several structured predictionapplications. While high-performance computing devices such asGPUs has made developing very powerful deep neural networkspossible, it is not feasible to run these networks on low-cost, lowpowercomputing devices such as embedded CPUs or even embeddedGPUs. As such, there has been a lot of recent interestto produce efficient deep neural network architectures that can berun on small computing devices. Motivated by this, the idea ofStochasticNets was introduced, where deep neural networks areformed by leveraging random graph theory. It has been shownthat StochasticNet can form new networks with 2X or 3X architecturalefficiency while maintaining modeling accuracy. Motivated bythese promising results, here we investigate the idea of Stochastic-Net in StochasticNet (SiS), where highly-efficient deep neural networkswith Network in Network (NiN) architectures are formed ina stochastic manner. Such networks have an intertwining structurecomposed of convolutional layers and micro neural networksto boost the modeling accuracy. The experimental results showthat SiS can form deep neural networks with NiN architectures thathave 4X greater architectural efficiency with only a 2% dropin accuracy for the CIFAR10 dataset. The results are even morepromising for the SVHN dataset, where SiS formed deep neuralnetworks with NiN architectures that have 11.5X greater architecturalefficiency with only a 1% decrease in modeling accuracy.

Partial turbulence simulation method for predicting peak wind loads on small structures and building appurtenances

Large-scale wind tunnel testing is preferred for small structures and building appurtenances for maintaining modeling accuracy and minimizing Reynolds number effects. In these circumstances the ability to obtain a large enough turbulence integral scale is usually compromised by the limited dimensions of the wind tunnel. So, it is not normally possible to fully simulate the low frequency end of the turbulence spectrum. In this paper the approach is taken of dividing the turbulence into two distinct statistical processes, one at high frequencies which can be simulated in the wind tunnel, and one at low frequencies which can be treated in post-test analysis using the assumptions of quasi-steady theory. In this Partial Turbulence Simulation (PTS) method the contribution of both the high and low frequency turbulence on the wind loads on structures is included by using the probability of load from each of the two processes, with one part coming from the wind tunnel data representing the high frequency component and the remainder from the assumed probability distribution (taken in this paper as Gaussian for generic boundary layer flow) of the missing low frequency component. The two processes are approximated as independent of each other. The efficacy and validity of the method and its various assumptions are assessed by comparing predicted local peak pressure coefficients from tests on large scale models of the Silsoe cube and Texas Tech University (TTU) building in the Wall of Wind facility at Florida International University (FIU) with the corresponding full-scale data. Generally good agreement was found between the model results and full scale, particularly when comparing the highest overall peak pressure coefficients. These results, although limited to peak local pressures on the two test buildings for which good full scale data are available, are encouraging and invite further experiments to explore the range of applicability of the PTS method. This method, although developed in the Wall of Wind facility at FIU, can be equally used in conventional boundary layer wind tunnels and has the potential to enhance the ability of existing boundary layer wind tunnel facilities to predict full scale wind loads.

Multiscale multiphysics and multidomain models--flexibility and rigidity.

The emerging complexity of large macromolecules has led to challenges in their full scale theoretical description and computer simulation. Multiscale multiphysics and multidomain models have been introduced to reduce the number of degrees of freedom while maintaining modeling accuracy and achieving computational efficiency. A total energy functional is constructed to put energies for polar and nonpolar solvation, chemical potential, fluid flow, molecular mechanics, and elastic dynamics on an equal footing. The variational principle is utilized to derive coupled governing equations for the above mentioned multiphysical descriptions. Among these governing equations is the Poisson-Boltzmann equation which describes continuum electrostatics with atomic charges. The present work introduces the theory of continuum elasticity with atomic rigidity (CEWAR). The essence of CEWAR is to formulate the shear modulus as a continuous function of atomic rigidity. As a result, the dynamics complexity of a macromolecular system is separated from its static complexity so that the more time-consuming dynamics is handled with continuum elasticity theory, while the less time-consuming static analysis is pursued with atomic approaches. We propose a simple method, flexibility-rigidity index (FRI), to analyze macromolecular flexibility and rigidity in atomic detail. The construction of FRI relies on the fundamental assumption that protein functions, such as flexibility, rigidity, and energy, are entirely determined by the structure of the protein and its environment, although the structure is in turn determined by all the interactions. As such, the FRI measures the topological connectivity of protein atoms or residues and characterizes the geometric compactness of the protein structure. As a consequence, the FRI does not resort to the interaction Hamiltonian and bypasses matrix diagonalization, which underpins most other flexibility analysis methods. FRI's computational complexity is of O(N(2)) at most, where N is the number of atoms or residues, in contrast to O(N(3)) for Hamiltonian based methods. We demonstrate that the proposed FRI gives rise to accurate prediction of protein B-Factor for a set of 263 proteins. We show that a parameter free FRI is able to achieve about 95% accuracy of the parameter optimized FRI. An interpolation algorithm is developed to construct continuous atomic flexibility functions for visualization and use with CEWAR.

Reduced-order discrete element method modeling

Reduced-order modeling (ROM) techniques are playing a very significant role in the recent literature bridging the gap between computationally expensive simulators and their application in optimization and control of distributed parameter systems. In modeling of solid-based processes, Discrete Element Method (DEM) is a very popular high-fidelity modeling approach which enables the tracking of discrete entities as they are processed inside complex geometries interacting with each other and with process boundaries. However, the computational cost of DEM models limits their applicability to simulation of specific and limited number of cases. In this work, we develop the DE–ROM (Discrete Element–Reducer-order model) approach which reduces the computational cost—while maintaining model accuracy—of expensive DEM models. The proposed methodology is based on Principal Component Analysis (PCA), combined with surrogate-model mapping and data imputation for the cases where there is missing information. The results of the application of the developed method to the powder mixing process, demonstrate that the DE–ROM model produces accurate profiles of distributed particle properties in real time. Finally, the developed model is used for the identification of the optimal design and operating conditions of the mixing process.

Use of upscaled elevation and surface roughness data in two-dimensional surface water models

In this paper, we present an approach that uses a combination of cell-block- and cell-face-averaging of high-resolution cell elevation and roughness data to upscale hydraulic parameters and accurately simulate surface water flow in relatively low-resolution numerical models. The method developed allows channelized features that preferentially connect large-scale grid cells at cell interfaces to be represented in models where these features are significantly smaller than the selected grid size. The developed upscaling approach has been implemented in a two-dimensional finite difference model that solves a diffusive wave approximation of the depth-integrated shallow surface water equations using preconditioned Newton–Krylov methods. Computational results are presented to show the effectiveness of the mixed cell-block and cell-face averaging upscaling approach in maintaining model accuracy, reducing model run-times, and how decreased grid resolution affects errors. Application examples demonstrate that sub-grid roughness coefficient variations have a larger effect on simulated error than sub-grid elevation variations.