Numerical Machine Learning Research Articles

Research of the numerical simulation and machine learning backpropagation networks analysis of the sound absorption properties of cellular soundproofing materials

Cellular soundproofing materials have been critical for lowering incidence pressure waves, which have led to their wide application in studios, auditoria, automotive and aerospace applications, among others. This work, therefore, premiers the utilisation of machine learning backpropagation algorithm of artificial neural networks (ANN) to develop and train a series of datasets of sound absorption properties (noise reduction coefficient, sound absorption average, and area under the sound absorption spectrum) obtained via numerical modelling and simulations. ANN models were established to analyze these sound absorption properties in relation to several non–acoustical parameters associated with cellular materials, including their thickness. Modelling and simulation revealed the relative importance of non–acoustic parameters on sound absorption spectra for these structures. ANN models predicted output signals that almost overlapped with their respective “real” signals with correlation coefficients over 95 %. Model confidence was proposed to analytically predict sound absorption phenomena accurately for cellular materials characterized by permeability ranging between 0.5–3.0 × 10−09 m2 or viscous characteristic lengths ranging between 80 and 300 μm. Analytical models based on optimised synaptic weights, biases, non–acoustic parameters, and material thickness could assist manufacturers design soundproofing materials for noise reduction and reverberation control.

Research on underwater acoustic field prediction method based on physics-informed neural network

In the field of underwater acoustic field prediction, numerical simulation methods and machine learning techniques are two commonly used methods. However, the numerical simulation method requires grid division. The machine learning method can only sometimes analyze the physical significance of the model. To address these problems, this paper proposes an underwater acoustic field prediction method based on a physics-informed neural network (UAFP-PINN). Firstly, a loss function incorporating physical constraints is introduced, incorporating the Helmholtz equation that describes the characteristics of the underwater acoustic field. This loss function is a foundation for establishing the underwater acoustic field prediction model using a physics-informed neural network. The model takes the coordinate information of the acoustic field point as input and employs a fully connected deep neural network to output the predicted values of the coordinates. The predicted value is refined using the loss function with physical information, ensuring the trained model possesses clear physical significance. Finally, the proposed prediction model is analyzed and validated in two dimensions: the two-dimensional acoustic field and the three-dimensional acoustic field. The results show that the mean square error between the prediction and simulation values of the two-dimensional model is only 0.01. The proposed model can effectively predict the distribution of the two-dimensional underwater sound field, and the model can also predict the sound field in the three-dimensional space.

Air quality analysis and modeling in urban area: A review study

Urban areas across the world are experiencing significant challenges related to air quality, with adverse effects on public health, the environment, and overall quality of life. This thesis aims to investigate and model air quality in urban environments, focusing on the identification of pollution sources and the evaluation of potential mitigation strategies. The research begins by analyzing the complex interplay of factors influencing urban air quality, including emissions from vehicular traffic, industrial activities, and meteorological conditions. Various modeling techniques, such as numerical simulations, statistical analyses, and machine learning algorithms, are employed to assess the spatiotemporal distribution of air pollutants within urban areas. The overview is carried out to focus on various parameters such as emission sources, meteorological parameters, and topographical features to predict pollutant concentrations. The study also explores the effectiveness of different mitigation strategies to improve urban air quality. These strategies may include urban planning measures, such as the design of green spaces and pedestrian-friendly zones, as well as transportation-related interventions like the promotion of electric vehicles and the enhancement of public transportation systems. The impact of these strategies on reducing pollutant concentrations is quantitatively assessed using the developed air quality models. Furthermore, the study considers the implications of climate change on urban air quality, acknowledging that changing climate patterns can exacerbate air pollution issues. Potential adaptation strategies to address these challenges are discussed. In conclusion, this thesis offers a comprehensive analysis of air quality modeling in urban areas, providing insights into the sources of pollution and the effectiveness of mitigation measures. The research aims to inform policymakers, urban planners, and environmental stakeholders in their efforts to create healthier and more sustainable urban environments.

Highly strain-rate sensitive and ductile composite materials combining soft with stiff TPMS polymer-based interpenetrating phases

The current work investigates the engineering of strain-rate sensitive and ductile composites through the combination of stiff, triply-periodic-minimal-surface (TPMS) with soft, rubber-type material phases. Different interpenetrating phase composite (IPC) designs are considered, with Diamond, Fisher-Koch, and IWP TPMS architectures as reinforcement phase topologies. The loading rate sensitivity of the arising IPCs is evaluated at different strain-rates, revealing an effective composite performance that well separates from the either brittle or substantially low-strength behavior of its constituent materials. Significant strain-rate effects are recorded, which relate to increased stiffness and energy absorption attributes upon substantial ductility, with the actual performance to depend on the underlying reinforcement phase topology. The effective strain-rate IPC constitutive response is well-captured by dedicated, low numerical cost and high accuracy machine learning models, whereas Digital Image Correlation (DIC) and finite element analysis are employed to provide insights in the inner strain and stress fields developed and failure modes observed. Overall, energy absorptions higher than 10 and up to 14 MJ m−3 are reported for strain-rates in the order of 100, furnishing specific energy absorptions (SEA) in the uppermost range of the effective composite material behaviors up to now reported, upon exceptionally high crushing force efficiencies (CFE) at moderate densities.

Progress and Challenges of Integrated Machine Learning and Traditional Numerical Algorithms: Taking Reservoir Numerical Simulation as an Example

Machine learning techniques have garnered significant attention in various engineering disciplines due to their potential and benefits. Specifically, in reservoir numerical simulations, the core process revolves around solving the partial differential equations delineating oil, gas, and water flow dynamics in porous media. Discretizing these partial differential equations via numerical methods is one cornerstone of this simulation process. The synergy between traditional numerical methods and machine learning can enhance the precision of partial differential equation discretization. Moreover, machine learning algorithms can be employed to solve partial differential equations directly, yielding rapid convergence, heightened computational efficiency, and accuracies surpassing 95%. This manuscript offers an overview of the predominant numerical methods in reservoir simulations, focusing on integrating machine learning methodologies. The innovations in fusing deep learning techniques to solve reservoir partial differential equations are illuminated, coupled with a concise discussion of their inherent advantages and constraints. As machine learning continues to evolve, its conjunction with numerical methods is poised to be pivotal in addressing complex reservoir engineering challenges.

Research on neutron diffusion equation and nuclear thermal coupling method based on gradient updating finite volume method

To solve the problem that the neutron diffusion equation has infinite solutions when the deterministic method is used to calculate the neutron diffusion equation, this paper uses the power calculation equation as a supplementary equation to form a differential and integral equation groups with the neutron diffusion equation, and designs a numerical analysis and machine learning coupling algorithm based on the gradient updating finite volume method that can be used to solve the above differential and integral equation groups. The algorithm decouples the multigroup neutron diffusion equations by source iteration. Then the finite volume method is used to solve the one group neutron diffusion equation and the reactor core power. Based on the power difference calculated by two adjacent iterations, the neutron flux density is updated by the gradient descent method until it converges. Finally, the algorithm can be used to calculate the neutron flux density and volumetric heat release rate of each energy group under the specified power. Taking 5 × 5 rod bundle channel as an example, the feasibility of the algorithm is verified in Fluent 18.0 by writing the UDF script. In addition, the thermal–hydraulic parameters such as temperature and velocity in the reactor can be further calculated according to the volumetric heat release rate, and a same set of grids can also be used to achieve the nuclear thermal coupling calculation. This algorithm can provide a new calculation method for the neutron diffusion equation, and also provide a theoretical basis for the development of application algorithms in the field of reactor physics using deep learning.

Comparative Analysis of a Numerical Method and Machine Learning Methods of Temperature Determination of a Doped Lubricating Layer with Experimental Data

Comparative Analysis of a Numerical Method and Machine Learning Methods of Temperature Determination of a Doped Lubricating Layer with Experimental Data

Data-Driven Framework toward Accurate Prediction of Interfacial Thermal Resistance in Particulate-Filled Composites.

The interfacial thermal resistance (ITR) inside the particulate-filled polymer composite is a bottleneck for improving the thermal conductivity (TC) of the material. Getting full knowledge of the ITR is crucial to the material design as well as to a faithful prediction of TC of the composite. However, a method fully taking into account the local circumstances inside the composite is yet to be developed to precisely characterize the ITR. Here, we propose a comprehensive framework combining high-throughput numerical simulations, machine learning and optimization algorithms, and experiments, which is demonstrated to be robust for the accurate determination of ITRs inside the particulate-filled composites. The strategy extracts as much information as possible about the structure and heat transfer characteristics of the composite based on simple experiments, which lays the foundation for the method to be effective. We show that the polymer-filler ITRs and the effective filler-filler contact ITRs predicted with the method faithfully represent the true characteristics inside the composite materials; they also provide the exact effective parameters, which cannot be obtained from experiments, for accurate numerical prediction of TCs of composite materials with high efficiency. As a result, the framework not only provides a robust tool for accurate characterization of ITRs inside composites but also paves the way for virtual high-throughput formula screening of thermally conductive composite materials that could be used in industrial product design.

Population-Level Immunity for Transient Suppression of COVID-19 Waves in Japan from April 2021 to September 2022.

Multiple COVID-19 waves have been observed worldwide, with varying numbers of positive cases. Population-level immunity can partly explain a transient suppression of epidemic waves, including immunity acquired after vaccination strategies. In this study, we aimed to estimate population-level immunity in 47 Japanese prefectures during the three waves from April 2021 to September 2022. For each wave, characterized by the predominant variants, namely, Delta, Omicron, and BA.5, the estimated rates of population-level immunity in the 10-64-years age group, wherein the most positive cases were observed, were 20%, 35%, and 45%, respectively. The number of infected cases in the BA.5 wave was inversely associated with the vaccination rates for the second and third injections. We employed machine learning to replicate positive cases in three Japanese prefectures to validate the reliability of our model for population-level immunity. Using interpolation based on machine learning, we estimated the impact of behavioral factors and vaccination on the fifth wave of new positive cases that occurred during the Tokyo 2020 Olympic Games. Our computational results highlighted the critical role of population-level immunity, such as vaccination, in infection suppression. These findings underscore the importance of estimating and monitoring population-level immunity to predict the number of infected cases in future waves. Such estimations that combine numerical derivation and machine learning are of utmost significance for effective management of medical resources, including the vaccination strategy.

Comparative Analysis of a Numerical Method and Machine Learning Methods of Temperature Determination of a Doped Lubricating Layer with Experimental Data

This article compares machine learning methods and a numerical method of determination of the doped lubricating layer with experimental data. Based on the sweep method, the one-dimensional Fourier heat equation with boundary and initial conditions is solved. As a result of comparing numerical and predictive data with experiments, it can be concluded that machine learning models are better at predicting results compared to numerical data

Predicting the particle-agglomeration effect on the equivalent mechanical properties of dispersion nuclear fuel by machine learning

The optimal design of dispersion nuclear fuel necessitates precise correlations between the particle distribution and mechanical properties, which can be established by combining the numerical method, data-driven technique, and machine learning model. In this study, we developed an automated high-throughput workflow to rapidly generate massive number of dispersion nuclear fuel meat models with different particle distributions, and further built the machine learning database for predicting the crucial mechanical properties. Effects of the particle-agglomeration behavior on mechanical properties of the dispersion fuel meat were analyzed in length. The automated workflow includes the entire parametric modeling, parallel computation and post-processing of simulated results, through which we can calculate the equivalent elastic modulus and the maximum Mises stress of the dispersed microstructure with randomly distributed fuel particles. The Fourier distribution function is utilized to characterize the random particle distribution configuration. The analysis suggests that the particle distribution configuration, the volume fraction and the number of agglomerate particles are the crucial factors determining the performance of dispersion nuclear fuel. Furthermore, through utilizing the database constructed by high-throughput computing, the Gaussian process regression algorithm was successfully applied to accurately forecast the mechanical properties in the dispersion fuel meat. This work lays a foundation for optimizing the design of high-performance dispersion nuclear fuel, quickly estimating mechanical properties of composite structures containing randomly dispersed particles, and further extending to analogous systems.

Modeling Hydrodynamic Behavior of the Ottawa River: Harnessing the Power of Numerical Simulation and Machine Learning for Enhanced Predictability

The Ottawa River Watershed is a vast area that stretches across Ontario and Quebec and holds great importance for Canada’s people, economy, and collective history, both in the present and the future. The river has faced numerous floods in recent years due to climate change. The most significant flood occurred in 2019, surpassing a 100-year flood event, and serves as a stark reminder of how climate change impacts our environment. Considering the limitations of machine learning (ML) models, which heavily rely on historical data used during training, they may struggle to accurately predict such “non-experienced” or “unseen” floods that were not encountered during the training process. To tackle this challenge, our study has utilized a combination of numerical modeling and ML to create an integrated methodology. Indeed, a comprehensive dataset of river flow discharge was generated using a numerical model, encompassing a wide range of potential future floods. This significantly improved the ML training process to generalize the accuracy of results. Utilizing this dataset, a novel ML model called the Expanded Framework of Group Method of Data Handling (EFGMDH) has been developed. Its purpose is to provide decision-makers with explicit equations for estimating three crucial hydrodynamic characteristics of the Ottawa River: floodplain width, flow velocity, and river flow depth. These predictions rely on various inputs, including the location of the desired cross-section, river slope, Manning roughness coefficient at different river sections (right, left, and middle), and river flow discharge. To establish practical models for each of the aforementioned hydrodynamic characteristics of the Ottawa River, different input combinations were tested to identify the most optimal ones. The EFGMDH model demonstrated high accuracy throughout the training and testing stages, achieving an R2 value exceeding 0.99. The proposed model’s exceptional performance demonstrates its reliability and practical applications for the study area.

Development of advanced hybrid mechanistic-artificial intelligence computational model for learning of numerical data of flow in porous membranes

Numerical analysis and machine learning regression were carried out for understanding and description of ozonation process combined with membrane separation. The main focus was on mass transfer modeling to simulate concentration distribution of ozone in liquid phase. Machine learning regression models were utilized to decrease the computational expenses of CFD (Computational Fluid Dynamics) simulations. For machine learning models, we compared the performance of three regression models, namely Gaussian Process Regression (GPR), Deep Neural Network (DNN), and Extreme Gradient Boosting (XGB or XGboost) for predicting the obtained output of CFD simulations which was ozone concentration in the feed, i.e., C (mol/m3) as function of radial and axial coordinates. Hyper-parameter tuning for these models was performed using the Tabu Search algorithm. The results indicate that both GPR and DNN models achieved excellent prediction accuracy, with R-squared values of 0.99999 and 0.99998, respectively. Also, the RMSE for GPR and DNN were 3.8481E-03 and 4.9835E-03, respectively, while their maximum errors were 6.08501E-02 and 4.47909E-02, respectively. On the other hand, the XGB model showed a lower performance, with an R-squared value of 0.99402, an RMSE of 8.9484E-02, and a maximum error of 4.51381E-01

Modified Taylor Impact Tests with Profiled Copper Cylinders: Experiment and Optimization of Dislocation Plasticity Model.

Current progress in numerical simulations and machine learning allows one to apply complex loading conditions for the identification of parameters in plasticity models. This possibility expands the spectrum of examined deformed states and makes the identified model more consistent with engineering practice. A combined experimental-numerical approach to identify the model parameters and study the dynamic plasticity of metals is developed and applied to the case of cold-rolled OFHC copper. In the experimental part, profiled projectiles (reduced cylinders or cones in the head part) are proposed for the Taylor impact problem for the first time for material characterization. These projectiles allow us to reach large plastic deformations with true strains up to 1.3 at strain rates up to 105 s-1 at impact velocities below 130 m/s. The experimental results are used for the optimization of parameters of the dislocation plasticity model implemented in 3D with the numerical scheme of smoothed particle hydrodynamics (SPH). A Bayesian statistical method in combination with a trained artificial neural network as an SPH emulator is applied to optimize the parameters of the dislocation plasticity model. It is shown that classical Taylor cylinders are not enough for a univocal selection of the model parameters, while the profiled cylinders provide better optimization even if used separately. The combination of different shapes and an increase in the number of experiments increase the quality of optimization. The optimized numerical model is successfully validated by the experimental data about the shock wave profiles in flyer plate experiments from the literature. In total, a cheap, simple, but efficient route for optimizing a dynamic plasticity model is proposed. The dislocation plasticity model is extended to estimate grain refinement and volume fractions of weakened areas in comparison with experimental observations.

Correcting Subseasonal Forecast Errors with an Explainable ANN to Understand Misrepresented Sources of Predictability of European Summer Temperatures

Abstract Subseasonal forecasts are challenging for numerical weather prediction (NWP) and machine learning models alike. Forecasting 2-m temperature (t2m) with a lead time of 2 or more weeks requires a forward model to integrate multiple complex interactions, like oceanic and land surface conditions leading to predictable weather patterns. NWP models represent these interactions imperfectly, meaning that in certain conditions, errors accumulate and model predictability deviates from real predictability, often for poorly understood reasons. To advance that understanding, this paper corrects conditional errors in NWP forecasts with an artificial neural network (ANN). The ANN postprocesses ECMWF extended-range summer temperature forecasts by learning to correct the ECMWF-predicted probability that monthly t2m in western and central Europe exceeds the climatological median. Predictors are objectively selected from ECMWF forecasts themselves, and from states at initialization, i.e., the ERA5 reanalysis. The latter allows the ANN to account for sources of predictability that are biased in the NWP model itself. We attribute ANN corrections with two explainable artificial intelligence (AI) tools. This reveals that certain erroneous forecasts relate to tropical western Pacific Ocean sea surface temperatures at initialization. We conjecture that the atmospheric teleconnection following this source of predictability is imperfectly represented by the ECMWF model. Correcting the associated conditional errors with the ANN improves forecast skill. Significance Statement We want to understand occasions in which a numerical weather prediction (NWP) model fails to forecast a predictable event existing in the real world. For forecasts of European summer weather more than 2 weeks in advance, real predictable events are rare. When misrepresented by the model, predicted future states become needlessly biased. We diagnose these missed opportunities with an explainable neural network. The neural network is aware of the initial state and learns to correct the NWP forecast on occasions when it misrepresents a teleconnection from the western tropical Pacific Ocean to Europe. The explainable architecture can be useful for other applications in which conditional model errors need to be understood and corrected.

Exploration of the mechanism and prediction of jet pump feature structure optimization for the enhanced pumping performance

It is widely concerned that the non-ideal design of characteristic parameters of throat pipe of jet pump leads to the low pumping efficiency. The paper studies the pumping efficiency of jet pump under different characteristic parameters through numerical calculation to determine the influences of single parameter on pumping efficiency. And the coupling influences of parameters are predicted through XGBoost technology. The research shows that the coupling relation between the pumping efficiency of jet pump and multiple characteristic parameters can be predicted with the accuracy of 98.83%. The reasonable choice of characteristic parameters can improve pumping efficiency by 23.6%. The research will help designers to look for the best coupling of jet pump parameters accurately through numerical calculation and machine learning technologies.

Speed estimation of a car at impact with a W-beam guardrail using numerical simulations and machine learning

This paper aimed at developing a new method of estimating the impact speed of a passenger car at the moment of a crash into a W-beam road safety barrier. The determination of such a speed based on the accident outcomes is demanding, because often there is no access to full accident data. However, accurate determination of the impact speed is one of the key elements in the reconstruction of road accidents. A machine learning algorithm was used to create the speed estimation model. The model was based on regression trees algorithms, with base regressors forming a final voting ensemble. The model was trained, validated, and tested using a database containing results from full-scale crash tests and numerical simulations. The developed machine learning model had a mean absolute error of 6.76 km/h with a standard deviation of 1.01 km/h on the cross-validation set, and a coefficient of determination, R2, of 0.85. This model was used to estimate the impact speed of the vehicle in three real road accidents with the W-beam barrier, and then the determined speeds were used in additional simulations to verify the results. A good quantitative and qualitative agreement between the simulation and accident outcomes was achieved, and this confirmed that the proposed method and the developed ML models combined with numerical simulations and full-scale crash tests can be effective tools for estimating the speed of the vehicle at impact with a roadside barrier.

Risk assessment models of power transmission lines undergoing heavy ice at mountain zones based on numerical model and machine learning

To deliver the cleaner energy, such as water, wind and solar power, from the remote mountainous areas to the power load areas, overhead power transmission lines are widely employed around the world. It is not an easy job to guarantee the safety of the long distance power transmission with hundreds or even thousands of kilometers of lines, especially crossing heavy ice zones at mountains. In this work, a numerical model to simulate the static and dynamic responses of transmission lines undergoing heavy ice is proposed and validated. Using the model, parameter study under different line sections depending on the mountain terrains, height difference and tension in conductors is carried out. It is obtained that the ice-induced dynamic characteristics of transmission lines at mountain zones are highly different with those of traditional design standard, with more complicated and dangerous. Therefore, to assess the risk level of transmission lines at mountain zones, the numerical model is used to gather the big dataset. Then, five machine learning-based models, including three tree-based and two artificial neural network (ANN)-based models, are developed and compared to find the optimal one. It is found that the multi-layer feed-forward deep ANN (MLF-DNN)-based model behaves optimal with the best evaluation scores. Based on the models, the predominant input parameters to affect the risk level are obtained and some suggestions to improve the ice resistance for the design and operation of transmission lines at mountain zones are proposed.

Online bearing fault diagnosis using numerical simulation models and machine learning classifications

Digital twin (DT) is the embodiment of the most advanced achievements of the current simulation technology theory development and the direction of intelligent development in the future. However, it is a great challenge to really integrate it into practical project application. Motivated by DT, an application method combining numerical simulation model and machine learning classification is proposed to show the advantages of digital twin. To ensure the reliability of the twin model, it is necessary to build a simulation model using a mature dynamic model, and modify it through the Pearson correlation coefficient (PCC) which is a kind of model online learning. Then, the required fault type is introduced by modifying the relevant fault influence factors, which is synchronously inserted into the normal operation model to obtain the normal, fault and other simulation numerical data. Finally, the machine learning model is used to predict the probability of each fault and feedback the impact value to the actual operation to guide the adjustment of actual parameters and the determination of maintenance plans. The experimental results show that this method can effectively predict the possibility of bearing failure synchronously and guide the adjustment and maintenance of actual bearing operating parameters.

Exploiting image logs to reduce drilling hazards: an innovative Artificial Intelligence methodology applied in East Africa

SUMMARY Image logs are crucial for petrophysical and structural characterization of drilled formations. Logging while drilling (LWD) image logs allow near real-time characterization of formations during drilling. Interpretation time required by humans can become critical, as structural interpretation provides the most crucial information for decision-making steps with respect to drilling operations. An example of this occurred in a deepwater system of reservoirs in East Africa. During the drilling of some wells, drilling was interrupted by unexpected borehole stability issues observed in a mass transport deposit. Those problems can detrimentally affect operations by impacting the efficiency of drilling. Therefore, a procedure for the automation of LWD data analysis was developed aimed at reducing drilling risk associated with borehole stability. A workflow is proposed based on an innovative assortment of AI methodologies for automatic real-time interpretation of LWD image log, applied to density image logs. The goal is to mimic manual human dip picking, combining computer vision, numerical series analysis and machine learning. Computer vision techniques are applied to detect the main density contrasts, representing the most likely surfaces. Sharp contrasts in the image are then investigated by dynamic time warping for similarity analysis. Finally, the geological plane is defined as a regression plane passing through the most similar contrasts, and a surface confidence criterion is defined based on the homogeneity of contrast along the curve suggested by the regression. Feature identification by a geologist was used as benchmark for comparison and metrics. Accuracy is referred to the difference between apparent dip and azimuth of geological surfaces coming from human and automated workflows, following the idea that two similar geometric results would eventually lead to equally similar geological considerations, and thus to similar operational measures. In general, a prediction accuracy of 60 per cent for the tested wells is observed, considering the total set of surfaces coming from the model. Nevertheless, the results suggest a positive correlation between the apparent dip of the discontinuities and the model error and a general less confidence for the surfaces with greater dip. By isolating the subset of low dipping features generated by the model, that is up to 35°, the accuracy appears to increase up to more than 70 per cent, suggesting that correlation is more difficult for high dipping geological planes. This class of curves also shows a general lower confidence and much of this latter class of surfaces are more likely to be found in the depth interval in which borehole stability issues occurred.