Back-propagation Artificial Neural Network Model Research Articles

Cadmium Elimination via Magnetic Biochar Derived from Cow Manure: Parameter Optimization and Mechanism Insights

Designing an efficient and recyclable adsorbent for cadmium pollution control is an urgent necessity. In this paper, cow manure, an abundant agricultural/animal husbandry byproduct, was employed as the raw material for the synthesis of magnetic cow manure biochar. The optimal preparation conditions were found using the response surface methodology model: 160 °C for the hydrothermal temperature, 600 °C for the pyrolysis temperature, and Fe-loading with 10 wt%. The optimal reaction conditions were also identified via the response surface methodology model: a dosage of 1 g·L−1, a pH of 7, and an initial concentration of 100 mg·L−1. The pseudo-second-order model and the Langmuir model were used to fit the Cd(II) adsorption, and the adsorption capacity was 612.43 mg·g−1. The adsorption was dominated by chemisorption with the mechanisms of ion-exchange, electrostatic attraction, pore-filling, co-precipitation, and the formation of complexations. Compared to the response surface methodology model, the back-propagation artificial neural network model fit the Cd(II) adsorption better as the error values were less. All these results demonstrate the potential application of CM for Cd(II) removal and its optimization through machine-learning processes.

Hybrid Artificial Neural Network-Based Models to Investigate Deformation Behavior of AZ31B Magnesium Alloy at Warm Tensile Deformation.

The uniaxial warm tensile experiments were carried out in deformation temperatures (50-250 °C) and strain rates (0.005 to 0.0167 s-1) to investigate the material workability and to predict flow stress of AZ31B magnesium alloy. The back-propagation artificial neural network (BP-ANN) model, a hybrid models with a genetic algorithm (GABP-ANN), and a constrained nonlinear function (CFBP-ANN) were investigated. In order to train the exploited machine learning models, the process parameters such as strain, strain rate, and temperature were accounted as inputs and flow stress was considered as output; moreover, the experimental flow stress values were also normalized to constructively run the neural networks and to achieve better generalization and stabilization in the trained network. Additionally, the proposed model's closeness and validness were quantified by coefficient of determination (R2), relative mean square error (RMSE), and average absolute relative error (AARE) metrics. The computed statistical outcomes disclose that the flow stress predicted by both GABP-ANN and CFBP-ANN models exhibited better closeness with the experimental data. Moreover, compared with the GABP-ANN model outcomes, the CFBP-ANN model has a relatively higher predictability. Thus, the outcomes confirm that the proposed CFBP-ANN model can result in the accurate description of AZ31 magnesium alloy deformation behavior, showing potential for the purpose of practicing finite element analysis.

Prediction of Soil Water Content Based on Hyperspectral Reflectance Combined with Competitive Adaptive Reweighted Sampling and Random Frog Feature Extraction and the Back-Propagation Artificial Neural Network Method

The soil water content (SWC) is a critical factor in agricultural production. To achieve real-time and nondestructive monitoring of the SWC, an experiment was conducted to measure the hyperspectral reflectance of soil samples with varying levels of water content. The soil samples were divided into two parts, SWC higher than field capacity (super-θf) and SWC lower than field capacity (sub-θf), and the outliers were detected by Monte Carlo cross-validation (MCCV). The raw spectra were processed using Savitzky–Golay (SG) smoothing and then the spectral feature variable of SWC was extracted by using a combination of competitive adaptive reweighted sampling (CARS) and random frog (Rfrog). Based on the extracted feature variables, an extreme learning machine (ELM), a back-propagation artificial neural network (BPANN), and a support vector machine (SVM) were used to establish the prediction model. The results showed that the accuracy of retrieving the SWC using the same model was poor, under two conditions, i.e., SWC above and below θf, mainly due to the influence of the lower accuracy of the super-θf part. The number of feature variables extracted by the sub-θf and super-θf datasets were 25 and 18, respectively, accounting for 1.85% and 1.33% of the raw spectra, and the variables were widely distributed in the NIR range. Among the models, the best results were achieved by the BPANN model for both the sub-θf and the super-θf datasets; the R2p, RMSEp, and RRMSE of the sub-θf samples were 0.941, 1.570%, and 6.685%, respectively. The R2p, RMSEp, and RRMSE of the super-θf samples were 0.764, 1.479%, and 4.205%, respectively. This study demonstrates that the CARS–Rfrog–BPANN method was reliable for the prediction of SWC.

Artificial intelligence optimization and controllable slow-release iron sulfide realizes efficient separation of copper and arsenic in strongly acidic wastewater

Iron sulfide (FeS) is a promising material for separating copper and arsenic from strongly acidic wastewater due to its S2− slow-release effect. However, uncertainties arise because of the constant changes in wastewater composition, affecting the selection of operating parameters and FeS types. In this study, the aging method was first used to prepare various controllable FeS nanoparticles to weaken the arsenic removal ability without affecting the copper removal. Orthogonal experiments were conducted, and the results identified the Cu/As ratio, H2SO4 concentration, and FeS dosage as the three main factors influencing the separation efficiency. The backpropagation artificial neural network (BP-ANN) model was established to determine the relationship between the influencing factors and the separation efficiency. The correlation coefficient (R) of overall model was 0.9923 after optimizing using genetic algorithm (GA). The BP-GA model was also solved using GA under specific constraints, predicting the best solution for the separation process in real-time. The predicted results show that the high temperature and long aging time of FeS were necessary to gain high separation efficiency, and the maximum separation factor can reached 1,400. This study provides a suitable sulfurizing material and a set of methods and models with robust flexibility that can successfully predict the separation efficiency of copper and arsenic from highly acidic environments.

Artificial neural network modelling for predicting classical air pressure profile curves in building drainage systems

Water trap seal loss can lead to the ingress of foul odours from the sewer network carrying harmful pathogens which can affect the health of building occupants. This loss is due to excessive negative or positive pressures within the system as a consequence of rapidly changing flow due to the unsteady nature of air and water flow discharges from appliances. This pressure regime within the system is often represented by a pressure profile along the height of the main vertical stack in the Building Drainage System (BDS). Experimental data from peer-reviewed literature and data from a unique 34-storey drainage test rig have been used as pressure profile data (Test data) for an Artificial Neural Network (ANN) model. Discharge input height (storey number) and discharge flow rate are considered to be the two independent input parameters and the pressure along the vertical stack is considered to be the output. In this work, both a Feed Forward and Back Propagation (FFBP) ANN model and a Radial Basis Function (RBF) ANN model have been used to train the algorithm. The work has confirmed the applicability of the FFBP-ANN model for steady two-phase fluid flow phenomena in BDS and allows for the prediction of pressures in a system for which no pressure data exists, by the prediction of modelled weights, based only on its physical and flow characteristics. Practical Application: Of great concern to designers of building drainage systems (BDS) is the control of pressure fluctuations within the system to prevent water trap seal loss. Prediction of a pressure profile based on the characteristics of the building – e.g. height, location of water discharge, discharge flow rate and ventilation type, would therefore increase confidence in designs, particularly in tall buildings. The work presented here addresses, for the first time, the applications of two ANN models for predicting the pressure profile in the BDS vertical stack of multi-storey buildings.

Revealing hot deformation behavior of inconel X-750 superalloy: A novel hot processing map coupled with grain boundary engineering

In this work, a series of isothermal hot compression tests were performed at the temperature of 900–1150 °C with the strain rate ranging from 0.01 to 10s−1 in Inconel X-750 alloy (referred as Alloy X-750). Based on the flow stress data, the back-propagation artificial neural network model was utilized to accurately predict the flow stress at elevated temperatures. Subsequently, by constructing a modified hot processing map coupled with the variation of Σ3n (n = 1, 2, and 3) twin boundaries fraction, the optimal processing parameters were determined to be in the range of 1100–1150 °C and 1-10s−1 for Alloy X-750. Moreover, the possibility of grain boundary engineering via hot deformation were also evaluated by using some critical quantification descriptors of twin related domains (TRDs), such as the number of grains in the TRDs, the lengths of the longest chain, and the average TRDs size. All the results can further verify the validity of modified hot processing map. Finally, the high-temperature flow behavior combined with the relevant microstructure characterization showed that discontinuous dynamic recrystallization is considered as the predominant softening mechanism, while the continuous dynamic recrystallization and particle-stimulated nucleation also additionally contribute to the recrystallization nucleation during the hot deformation of Alloy X-750.

Prediction of Spatial Peak Overpressure Profile of Air Blast Shocks Using Multiple Linear Regression and Artificial Neural Network

The shock peak overpressure yielded by energetic materials in the near-earth explosion causes damage to people and property. In the literature, several works have reported the effectiveness in obtaining the peak overpressure in the near field by numerical simulation, real tests, and empirical calculations. However, several issues are associated with the above methods, such as the time-consuming of a complex simulation, the difficulties in placing sensors in an obstacle block site, and the limited scope of applications for empirical equations. At present, there is no single model that can rapidly predict the spatial distribution of shock peak overpressure with ground reflection. In this regard, this paper aims to solve the problem of shock propagation in a large-scale space and obtain a model that can predict a wide range of spatial distribution of peak overpressure in the near-earth air blast simply and quickly. This work uses AUTODYN-2D to obtain the peak overpressure of the cylindrical charge explosion in the air near earth. The peak overpressure from the numerical models agrees with the empirical equations and the real blast results. On this basis, the multiple linear regression (MLR) and the optimal error back-propagation artificial neural network (ANN) models are tried and constructed with three geometric parameters as the input under a typical environment. The input parameters are the straight line scaled distance (SLSD), the vertical scaled distance (VSD), and the scaled height of burst (SHOB), respectively. The prediction results of the MLR and ANN models are verified and compared on the test set. Results show that the ANN model with the geometric parameters as the input is better than the MLR model in predicting the air blast peak overpressure. The former takes no more than 0.2s to compute 1441 sets of inputs, which provides a quick reference for predicting spatial peak overpressure in a large-scale shock wave of the air blast.

Generalized model for predicting the thermal conductivity of fine-grained soils

As a renewable energy, geothermal is considered as one of the important approaches to mitigate the energy crisis and climate change. The thermal conductivity of soils is a key parameter required in the utilization of geothermal resources, and it plays an important role in engineering, environmental science, earth science, and agriculture. Generally, the existing models that predict the thermal conductivity of soil are limited to specific types of soil and specific conditions. As a result, it is difficult to apply these models to engineering practice. Therefore, a simple, efficient, and generalized model capable of predicting the thermal conductivity of soils is necessary. In this work, a generalized model for calculating the thermal conductivity of fine-grained soils is established. This model considers the initial dry density, initial water content, temperature, and mineral composition. The proposed generalized model is based on a large amount of measured data from different countries and regions. Its validity is verified after being cross-validated with a back-propagation artificial neural network (BP-ANN) model. Compared with two existing models, the proposed model achieves the best thermal conductivity estimation of fine-grained soils. Additionally, it has a wide range of applications, requires only a few parameters, and does not need recalibration.

Modeling the Rheological Behavior of a Novel Hot Isostatic Pressed Powder Metallurgy Superalloy

The rheological behavior and deformation mechanisms of a new powder metallurgy (P/M) superalloy at various deformation conditions are researched. The deformation conditions have significant influence on the rheological stress. Increasing the deformation temperature or decreasing the strain rate can decrease the rheological stress. The discontinuous hardening and softening phenomena are observed at the strain rate of 1 s−1, resulting from the complex phase transformation and dynamic recrystallization. Besides, the deformation activity energy (Q) declines with increasing the strain. The phenomenon is attributed to the spheroidization of γ′ phase and the decreased content/aspect ratio of γ′ phase. The deformation mechanisms of the researched superalloy are the accumulation of dislocation, stacking faults shearing, dislocations pinned by γ′ phase, and the formation of microtwins during hot deformation. The strain‐compensated Arrhenius and particle swarm optimization‐based backpropagation artificial neural network (PSO‐BP ANN) models are established to predict the rheological behavior. Compared to the strain‐compensated Arrhenius equation, the developed PSO‐BP ANN model presents the higher accuracy in predicting the rheological behavior of the researched alloy. Furthermore, for the developed PSO‐BP ANN model, the correlation coefficient is 0.9995, and the root mean square error is 1.224 MPa. So, the forecasted rheological stresses are consistent with the measured ones.

Prediction of Regional Forest Biomass Using Machine Learning: A Case Study of Beijing, China

Dynamic changes in forest biomass are closely related to the carbon cycle, climate change, forest productivity and biodiversity. However, most previous studies mainly focused on the calculation of current forest biomass, and only a few studies attempted to predict future dynamic changes in forest biomass which obtained uncertain results. Therefore, this study comprehensively considered the effects of multi-stage continuous survey data of forest permanent sample plots, site condition factors and corresponding meteorological factors using Beijing as an example. The geographic detector method was used to screen the key interfering factors that affect the growth of forest biomass. Then, based on the back-propagation artificial neural network (BP-ANN) and support vector machine (SVM) learning methods, 80% of the sample data were extracted to train the model, and thereby verify the prediction accuracy of different modeling methods using different training samples. The results showed that the forest biomass prediction models based on both the machine learning algorithms had good fitting accuracy, and there was no significant difference in the prediction results between the two models. However, the SVM model was better than the BP-ANN. While the BP-ANN model provided more volatile predictions, and the accuracy was above 80%, the prediction results of the SVM model were relatively stable, and the accuracy was above 90%. This study not only provides good technical support for the scientific estimation of regional forest biomass in the future, but also offers reliable basic data for sustainable forest management, planning decisions, forest carbon sequestration and sustainable development.

Predicting indoor particle concentration in mechanically ventilated classrooms using neural networks: Model development and generalization ability analysis

This study presents a comprehensive analysis of the model development (structure, parameter settings, and prediction accuracy) and generalization ability of neural networks in the classroom setting. A normalized database for model prediction and verification including indoor parameters from 27 high school classrooms over the course of four semesters was set up (a total of 14,112 sets of data for each variable). Three different models were compared: a back propagation artificial neural network (BP-ANN), a nonlinear autoregressive exogenous neural network (NARX), and a long-term and short-term memory neural network (LSTM). For the prediction of PM1 and PM2.5, which are impacted by outdoor PM concentrations, the best performance was achieved by the NARX model: the mean absolute percentage error (MAPE), root mean squared error (RMSE), and coefficient of determination (R2) were 41–55%, 0.45–1.27 μg/m3, and 0.81–0.87, respectively. The R2 of this model was 22% higher than that of the BP-ANN model and much higher than that of the LSTM model. For the prediction of indoor PM10, which are mostly emitted by indoor sources, relatively good performance was achieved by the LSTM model. The trained NARX model was used successfully to predict indoor time-series PM2.5 concentration and source strength in classrooms in five Chinese cities and five American cities. The calculated indoor source strength was 95.5–119.1 μg/h/P, within the range of the measured indoor source strength. The indoor PM2.5 concentrations calculated by a theoretical equation without the consideration of indoor source were 20.8–60.0% smaller than those calculated by the trained model.

Non-destructive determination of soluble solids content in intact apples using a self-made portable NIR diffuse reflectance instrument

The soluble solids content (SSC) is an important internal quality parameter of fruits that is monitored on the fruit market to meet different consumer demands. The aim of this study was to develop a self-made portable near-infrared (NIR) diffuse reflectance instrument to evaluate and monitor the SSC of intact apples. The visible and near-infrared (Vis-NIR) spectra of 118 ‘Fuji’ apples were collected using a Vis-NIR diffuse reflectance spectroscopic measurement system in the spectral range of 450–1100 nm. Savitzky-Golay convolution smoothing and first derivative methods were subsequently applied to eliminate noise interference and baseline shift. Successive projections algorithm (SPA) was performed to extract feature wavelengths from the pretreatment spectra, and the back-propagation artificial neural network (BP-ANN) and multivariate nonlinear regression (MNLR) models were employed to evaluate the predictive ability of the extracted and selected feature wavelengths for the SSC of apples. Seven wavelengths (881, 890, 901, 926, 941, 951, 978 nm) were selected as optimal feature wavelengths from the extracted feature wavelengths (18 wavelengths). The MNLR model (R2 = 0.953, RMSE = 0.391 %) exhibited a high prediction accuracy compared to the BP-ANN model (R2 = 0.865, RMSE = 0.754 %). Based on the results, NIR combined with the MNLR model could be applied in the self-made portable instrument to innovatively monitor the SSC of apples. Finally, a self-made portable NIR diffuse reflectance instrument was designed and developed, and the spectral information of 298 ‘Fuji’ apples was collected using the instrument. According to the results, the MNLR model could well predict the SSC of apples (R2 = 0.871, RMSE = 0.687 %). The overall results also revealed that the developed portable NIR instrument is a promising device for rapid non-destructive monitoring of fruit SSC, thereby meeting the practical application requirement of postharvest commercial processing systems. This instrument could also be beneficial to customers, growers, and producers.

Application of colorimetric sensor array combined with visible near-infrared spectroscopy for the matcha classification

Matcha tea powder is considered as an integral part of a healthy diet due to its enormous health benefits. In the current study, visible near-infrared (Vis-NIR) and colorimetric sensor array (CSA) techniques are applied to identify the matcha grades. The color-sensitive dyes reacted with their volatile compounds and the information was recorded by Vis-NIR spectroscopy, namely Vis-NIR-CSA. Specifically, three linear and three nonlinear classification models were compared, yielding the optimal identification rate by the back-propagation artificial neural network (BPANN) model with 99% and 98% in the calibration and prediction sets, respectively. The results indicated the sensor combined with the BPANN model could be applied satisfactorily in identification of different matcha grades. Additionally, the variations in volatile compounds between different matcha grades and eight characteristic volatile compounds were identified, which verified the sensor identification results. This study provided a scientific and novel method for the stability of matcha quality in production.

Artificial Neural Network Model for Predicting the Energy Loss of Natural Rubber Foam

In this work, a back-propagation artificial neural network model was optimally developed based on 25 experimental datasets for predicting the energy loss percentage of natural rubber foam. The foam specimens were prepared in a Banbury internal mixer at various conditions of mixing temperature (40-80°C), rotor speed (40-80 rpm), and mastication time (1-5 min). Stress-strain loops were analyzed by applying compressive force at aspeed test of 500 mm/min with the capacity of load cell 2.5 kN, the energy loss was further calculated. In model development, the experimental datasets were randomly divided into 70:15:15 for training, validation, and testing respectively. Levenberg-Marquardt algorithm was used as a training function was used because of its fast convergence. The prediction results revealed that the average prediction accuracy of the three models is higher than 90%. From a material design point of view, the developed model could be implemented to find the proper mixing conditions to obtain the material with the maximum energy dissipation.

Machine Learning for Design of Phosphorene Nanozyme Sensor and Its Intelligent Analysis of Clenbuterol in Animal-Derived Agro-Products

Extraordinary electronic performance and unique structural characteristic of black phosphorene (BP) often is used as electrode modified materials in electrochemical sensors. In this paper, a machine learning (ML) strategy for phosphorene nanozyme sensor and its the intelligent of clenbuterol (CLB) in pork and pig serum samples is prepared. The silver nanoparticles decorate BP to prevent oxidative degradation of BP surface and further hybridize with multi-walled carbon nanotubes (MWCNTs) composites containing nafion (Nf) treated with isopropanol (IP) to improve environmental stability and electrocatalytic capacity of BP. Back-propagation artificial neural network (BP-ANN) model combined with genetic algorithm (GA) is employed to optimize sensor parameters such as BP concentrations, MWCNTs concentrations and ratio of VNf:VIP, and compared with orthogonal experimental design (OED). Least square support vector machine, radial basis function and extreme learning machine are implemented to establish quantitative analysis model for CLB. The results showed that the CLB response current of BP sensor by BP-ANN-GA was improved 9.02% over OED method. Compared with the traditional linear regression, three models displayed better predictive performance, and LS-SVM was the best with the R 2 , RMSE and MAE and RPD of 0.9977, 0.0303, 0.0225, and 18.74, respectively. The average recoveries of CLB in pork and pig serum was 98.66% ∼ 101.67%, and its relative standard deviations was 0.19% ∼ 0.84%, indicating that electrochemical sensor using machine learning for intelligent analysis of CLB in animal-derived agro-products products was both feasible and practical.

Non-uniform Corrosion Mechanism and residual life forecast of marine engineering concrete reinforcement

Reinforced concrete structures are commonly employed in civil engineering due to their low cost and high performance. However, their exposure to the marine environment, which is typically characterized by elevated salt content and humidity levels, heightens the probability of corrosion. Such corrosion can lead to a decrease in the service life of these structures, which is of utmost importance to predict in order to prevent potential safety hazards. To address this issue, this study investigated the corrosion mechanism of reinforced concrete structures in the marine environment and found Cl- concentration to be the most critical parameter for the corrosion of steel. Afterward, a predictive model is proposed based on the Back Propagation Artificial Neural Network (BPANN) to estimate the residual service life of reinforced concrete structures in different marine environments by forecasting the Cl- concentration in reinforced concrete components. The results demonstrate the high accuracy of the proposed BPANN model, with a maximum relative error of 6.292 %. Consequently, the residual service life forecast system of reinforced concrete structures proposed in this study can provide useful guidance based on the simplistic framework for designing concrete structures in the marine environment by predicting their effective service life. Thus, this study provides valuable insights for promoting the safety and economic benefits of reinforced concrete structures in the marine environment.

Experimental data-driven and phenomenological modeling approaches targeting the enhancement of CaTiO3 photocatalytic efficiency

Optimization based on mathematical models has received growing attention in materials science. The first part of the work aims to optimize the photocatalytic activity of CaTiO3 for rhodamine B (RhB) degradation under UV-A irradiation, based on two developed mathematical models. Thirty hydrothermal syntheses of CaTiO3 were carried out according to the Box-Behnken design, considering synthesis temperature (X1), duration (X2), and concentration of shaping agent (X3) as input variables for two different Ca2+ sources: Ca(NO3)2 and CaCl2 (X4). The conversion of the studied pollutant after 4 h was situated in the range of 20–80%. Second-order regression and feedforward backpropagation artificial neural network models were developed, considering the synthesis conditions (X1, X2, X3, X4) as input and the conversion as output variables. The proposed model-based methodology for the optimization of CaTiO3 photocatalytic efficiency finally directed to the experimentally attained value of 96% for 200 °C (X1, opt), 23.17 h (X2, opt), 0.67 M (X3, opt), CaCl2 (X4, opt). Furthermore, in the second part of the study, the morphological, structural, textural, and optical properties of selected CaTiO3 samples were investigated via scanning electron microscopy, X-ray diffractometry, N2 sorption, and diffuse reflectance spectroscopy. Finally, the kinetic parameters for adsorption (kads: 0.10–0.67 m·h−1), desorption (kdes: 79–150 mmol·m−2·h−1), degradation (kdegr: 0.001–0.010 mmol·m−2(1−α)·W−α·h−1), and intensity exponent (α: 0.54–0.55) were fitted using an optimization procedure, considering the experimentally determined and model-predicted apparent reaction rate constants. The obtained kinetic parameters were correlated with the specific surface area of the catalysts and the conversion of RhB.

Predicting seismic interaction effect between soil and structure group using convolutional neural network

Quantifying the influence of seismic interaction between soil and structure group (SSGI) is of great significance to seismic design, retrofit, and damage assessment of structures in densely built urban areas. To this end, this study proposes a one-dimensional convolutional neural network (1D-CNN) model to rapidly predict the influence of adjacent structures on the maximum inter-story drifts and base shears of RC frame structures. Based on an experimentally validated three-dimensional finite element method, 890 pairs of soil-single structure versus soil-structure group systems under different earthquake loadings are simulated. The dataset comprising 890 groups of input (i.e., soil and structure group parameters, ground motion acceleration) and output data (i.e., changes in maximum inter-story drift and base shear) is constructed to train the machine learning model. Subsequently, sensitivity analysis is performed to identify optimal hyperparameters for training the 1D-CNN model, whereas a back propagation artificial neural network (BP-ANN) model is established to compare the model performance. Results indicate that compared with the BP-ANN model, the 1D-CNN model has a more stable and robust architecture and features superior prediction accuracy. In particular, the developed 1D-CNN model has a mean absolute error of less than 2.3% and an absolute error of less than 5.4% for 90% of cases in the testing set. The superior performance of the 1D-CNN model makes it an effective and efficient tool to be applied to predict the seismic responses of RC frame buildings under the SSGI effect.

Estimation of reference evapotranspiration using artificial neural network models for semi-arid region of Haryana

The study was conducted to evaluate performance of artificial neural network (ANN) models for estimating reference evapotranspiration (ET0) for semi-arid region of Haryana state. Ten years (2011-2020) daily weather data of maximum and minimum temperature, relative humidity, wind speed and sun shine hours was collected from the meteorological observatory at CCS HAU, Hisar. Multilayer perceptron feed forward back propagation ANN models were evaluated for different training algorithms (10), number of hidden layers (1-3) and number of neurons in hidden layers (1-30). Training algorithms compared in the study were heuristic techniques (GDA, GDX, RP), conjugate gradient (CGF, CGP, CGB, SCG), quasi-Newton (BFG, OSS) and Levenberg-Marquardt (LM). Results were compared against standard FAO Penman-Monteith method. The study revealed that best performance for ANN was found with LM algorithm in single layer of 13 neurons exhibiting RMSE, R, ME and RPD values 0.306, 0.986, 0.976 and 6.63, respectively. ANN models showed good performance in prediction of reference evapotranspiration.

Static Globularization Behavior and Artificial Neural Network Modeling during Post-Annealing of Wedge-Shaped Hot-Rolled Ti-55511 Alloy.

The globularization of the lamellar α phase by thermomechanical processing and subsequent annealing contributes to achieving the well-balanced strength and plasticity of titanium alloys. A high-throughput experimental method, wedge-shaped hot-rolling, was designed to obtain samples with gradient true strain distribution of 0~1.10. The samples with gradient strain distribution were annealed to obtain the gradient distribution of globularized α phase, which could rapidly assess the globularization fraction of α phase under different conditions. The static globularization behavior under various parameters was systematically studied. The applied prestrain provided the necessary driving force for static globularization during annealing. The substructure evolution and the boundary splitting occurred mainly at the early stage of annealing. The termination migration and the Ostwald ripening were dominant in the prolonged annealing. A backpropagation artificial neural network (BP-ANN) model for static globularization was developed, which coupled the factors of prestrain, annealing temperature, and annealing time. The average absolute relative errors (AARE) for the training and validation set are 3.17% and 3.22%, respectively. Further sensitivity analysis of the factors shows that the order of relative importance for static globularization is annealing temperature, prestrain and annealing time. The developed BP-ANN can precisely predict the static globularization kinetic curves without overfitting.