Integrating machine learning and finite element simulation for interpretable prediction of 3D-printed bone scaffold mechanics

In this paper, research work involved in the prediction of life of deep drawing die using artificial neural network (ANN) is described. The parameters affecting life of deep drawing die such as size and material of die block and punches are investigated through finite element (FE) analysis and the critical simulation values are determined. Based on FE analysis results, stress amplitude (S) vs. cycles to failure (N) approach is used for prediction of number of cycles of deep drawing die. The number of cycles gives the number of sheet metal parts that can be produced with the deep drawing die before its failure. The ANN model of the proposed system is developed using MATLAB. The data required for ANN are obtained from FE analysis. The generated output data from FE analysis are used to train ANN model. The developed ANN model predicts the life of deep drawing die in terms of number of sheet metal parts. Usefulness of the proposed system is demonstrated through an example of an industrial sheet metal part.

In this work, the electrical potential (EP) technique with an artificial neural networks (ANNs) for monitoring of nanostructures are used for the first time. This study employs an expert system to identify size and localize hidden nano-delamination (N.Del) inside layers of nano-pipe (N.P) manufactured from Basalt Fiber Reinforced Polymer (BFRP) laminate composite by using low-cost monitoring method of electrical potential (EP) technique with an artificial neural networks (ANNs), which are combined to decrease detection effort to discern N.Del location/size inside the N.P layers, with high accuracy, simple and low-cost. The dielectric properties of the N.P material are measured before and after N.Del introduced using arrays of electrical contacts and the variation in capacitance values, capacitance change and node potential distribution are analyzed. Using these changes in electrical potential due to N.Del, a finite element (FE) simulation model for N.Del location/size detection is generated by ANSYS and MATLAB, which are combined to simulate sensor characteristic, therefore, FE analyses are employed to make sets of data for the learning of the ANNs. The method is applied for the N.Del monitoring, to minimize the number of FE analysis in order to keep the cost and save the time of the assessment to a minimum. The FE results are in excellent agreement with an ANN and the experimental results available in the literature, thus validating the accuracy and reliability of the proposed technique.

In this paper, artificial neural networks (ANNs) are introduced to obtain improved regional low‐flow estimates at ungauged sites. A multilayer perceptron (MLP) network is used to identify the functional relationship between low‐flow quantiles and the physiographic variables. Each ANN is trained using the Levenberg‐Marquardt algorithm. To improve the generalization ability of a single ANN, several ANNs trained for the same task are used as an ensemble. The bootstrap aggregation (or bagging) approach is used to generate individual networks in the ensemble. The stacked generalization (or stacking) technique is adopted to combine the member networks of an ANN ensemble. The proposed approaches are applied to selected catchments in the province of Quebec, Canada, to obtain estimates for several representative low‐flow quantiles of summer and winter seasons. The jackknife validation procedure is used to evaluate the performance of the proposed models. The ANN‐based approaches are compared with the traditional parametric regression models. The results indicate that both the single and ensemble ANN models provide superior estimates than the traditional regression models. The ANN ensemble approaches provide better generalization ability than the single ANN models.

Horizontal openings in reinforced concrete (RC) beams are quite often used to accommodate service pipelines. Several research papers are available in the literature describing their effect. RC beams with vertical openings are commonly used to accommodate service lines in residential buildings in Kuwait. However, there are lack of design guidelines and best practices reported in the literature for RC beams with vertical openings, whereas the detailed guidelines are available for beams with horizontal openings. In the present paper, laboratory experiments are conducted on nine RC beams with and without vertical openings. Parametric study has been carried out using nonlinear finite element analysis (FEA) with changes in the diameter of the opening, various positions of the opening along the length and width of the beam, edge distance, etc. 50 finite element simulations were conducted. The FEA results are verified using the results from the laboratory experiments. The study showed that the load carrying capacity of the beam is reduced by 20% for the RC beam with vertical openings placed near the center of the beam compared to a solid beam without an opening. Significant reduction in load carrying capacity is observed for beams with an opening near the support (≈15%). The overall stiffness of the beam, crack pattern and failure modes were not affected due to the presence of the vertical opening. Furthermore, an artificial neural network (ANN) analysis is carried out using the FEA generated data. The results and observations from the ANN and FEA are in good agreement with experimental results.

To overcome the complication of jetty pile design process, artificial neural networks (ANN) are adopted. To generate the training samples for training ANN, finite element (FE) analysis was performed 50 times for 50 different design cases. The trained ANN was verified with another FE analysis case and then used as a structural analyzer. The multilayer neural network (MBPNN) with two hidden layers was used for ANN. The framework of MBPNN was defined as the input with the lateral forces on the jetty structure and the type of piles and the output with the stress ratio of the piles. The results from the MBPNN agree well with those from FE analysis. Particularly for more complex modes with hundreds of different design cases, the MBPNN would possibly substitute parametric studies with FE analysis saving design time and cost.

Elastography comprises a set of modalities that image the biomechanical properties of soft tissues for disease detection and diagnosis. Quasi-static ultrasound elastography, in particular, tracks sub-surface displacements resulting from an applied surface force. The local displacement information and measured surface loads may be used to compute a parametric summary of biomechanical properties; however, the inverse problem is under- determined, limiting most techniques to estimating a single linear-elastic parameter. We previously described a new method to develop mechanical models using a combination of computational mechanics and machine learning that circumvents the limitations associated with the inverse problem. The Autoprogressive method weaves together finite element analysis and artificial neural networks (ANNs) to develop empirical models of mechanical behavior using only measured force-displacement data. We are extending that work by incorporating spatial information with the material properties. Previously, the ANNs accepted only a strain vector input and computed the corresponding stress, meaning any spatial information was encoded in the finite element mesh. Now, using a pair of ANNs working in tandem with spatial coordinates included as part of the input, these new Cartesian ANNs are able to learn the spatially varying mechanical behavior of complex media. We show that a single Cartesian ANN is able to describe the same mechanical behavior of an object that previously required at least two ANNs. Furthermore, we show the new ANNs can learn complex material property distributions and reconstruct images of the Young’s modulus distribution, not merely classify, filter, or otherwise process an existing image. For the first time, we present results using Cartesian neural networks within the Autoprogressive Method to form elastic modulus images.

In this study, a hybridized neuro-genetic optimization methodology realized by embedding finite element analysis (FEA) trained artificial neural networks (ANN) into genetic algorithms (GA), is used to optimize temperature control in a ceramic based continuous flow polymerase chain reaction (CPCR) device. The CPCR device requires three thermally isolated reaction zones of 94 degrees C, 65 degrees C, and 72 degrees C for the denaturing, annealing, and extension processes, respectively, to complete a cycle of polymerase chain reaction. The most important aspect of temperature control in the CPCR is to maintain temperature distribution at each reaction zone with a precision of +/-1 degree C or better, irrespective of changing ambient conditions. Results obtained from the FEA simulation shows good comparison with published experimental work for the temperature control in each reaction zone of the microfluidic channels. The simulation data are then used to train the ANN to predict the temperature distribution of the microfluidic channel for various heater input power and fluid flow rate. Once trained, the ANN analysis is able to predict the temperature distribution in the microchannel in less than 20 min, whereas the FEA simulation takes approximately 7 h to do so. The final optimization of temperature control in the CPCR device is achieved by embedding the trained ANN results as a fitness function into GA. Finally, the GA optimized results are used to build a new FEA model for numerical simulation analysis. The simulation results for the neuro-genetic optimized CPCR model and the initial CPCR model are then compared. The neuro-genetic optimized model shows a significant improvement from the initial model, establishing the optimization method's superiority.

Bolted flange connections are one of the most commonly used joint types in aircraft structures. Typically, bolted flange connections are used in aircraft engines. The main duty of a bolted flange connection in an aircraft engine is to serve as the load transfer interface from one part of the engine to the other part of the engine. In aircraft structures, weight is a very critical parameter which has to be minimized while having the required margin of safety for the structural integrity. Therefore, optimum design of the bolted flange connection is crucial to minimize the weight. In the preliminary design stage of the bolted flange connection, many repetitive analyses have to be made in order to decide on the optimum design parameters of the bolted flange connection. Two main methods used for analyzing bolted flange connections are the hand calculations based on simplified approaches and finite element analysis (FEA). While hand calculations lack achieving optimum weight as they tend to give over safe results, finite element analysis is computationally expensive because of the non-linear feature of the problem due to contact definitions between the mating parts. In this study, a fast but very accurate design tool based on artificial neural network (ANN) is developed for the cylindrical bolted flange connection of a typical aircraft engine under combined axial and bending moment load. ANN uses the FEA database generated by taking permutations of the parametric design variables of the bolted flange connection. The selected parameters are the number of bolts, the bolt size, the flange thickness, the web thickness, the preload level of the bolt and the external combined loads of bending moment and axial force. The bolt reaction force and the average flange stress are taken as the output variables and the results of 12000 different finite element analyses are gathered to form a database for the training of the ANN. Results of the trained ANN are then compared with the finite element analysis results and it is shown that an excellent agreement exists between the ANN and the non-linear finite element analysis results within the training limits of the artificial neural network. We believe that the ANN established can be a very robust and accurate approximate model replacing the non-linear finite element solver in the optimization of the bolted flange connection of the aircraft engine to achieve weight reduction.

This study proposes a fast running model that interconnects input and output data for a single-pass cold bar drawing process through the use of Artificial Neural Network (ANN) and automatically generated a large volume of elastic-plastic finite element (FE) analysis results. The prediction accuracy of the FE analysis was verified by comparing the FE analysis with measurements from a drawing experiment. A Python-based script that automatically controls ABAQUS was coded to sequentially produce output data that varies according to the input data, which is a combination of 18 grades of steel and 1,000 process conditions. The ANN was trained using input and output data, and then a nine-dimensional fast running model was developed. The fast running model predicted the values of output variables (drawing force, strain at the center, strain on the surface, accumulated damage at the center, contact pressure, and the fracture (or non-fracture) of the material) in 0.1 second no matter how the mechanical properties of the steels and process conditions change. With this fast running model, engineers in the drawing industry can easily determine or modify the process conditions to improve productivity and product quality even when a grade of steel that has never been employed before is drawn.

A new method, termed simulated micromechanical models using artificial neural networks (MMANN), is proposed to generate micromechanical material models for nonlinear and damage behavior of heterogeneous materials. Artificial neural networks (ANN) are trained with results from detailed nonlinear finite-element (FE) analyses of a repeating unit cell (UC), with and without induced damage, e.g., voids or cracks between the fiber and matrix phases. The FE simulations are used to form the effective stress-strain response for a unit cell with different geometry and damage parameters. The FE analyses are performed for a relatively small number of applied strain paths and damage parameters. It is shown that MMANN material models of this type exhibit many interesting features, including different tension and compression response, that are usually difficult to model by conventional micromechanical approaches. MMANN material models can be easily applied in a displacement-based FE for nonlinear analysis of composite structures. Application examples are shown where micromodels are generated to represent the homogenized nonlinear multiaxial response of a unidirectional composite with and without damage. In the case of analysis with damage growth, thermodynamics with irreversible processes (TIP) is used to derive the response of an equivalent homogenized damage medium with evolution equations for damage. The proposed damage formulation incorporates the generalizations generated by the MMANN method for stresses and other possible responses from analysis results of unit cells with fixed levels of damage.

Stress concentration factor (SCF) is usually used to estimate the fatigue life of an offshore joint. Historically, parametric equations were used to estimate SCF based on a statistical analysis of experimental and finite element analysis (FEA) results, to reduce cost and time. These equations give the SCF at the saddle/crown position for simple joints and basic load cases. However, for modified or defective joints, the location of the maximum SCF can change. In such circumstances, the single-point SCF equation cannot be used to estimate the maximum value of SCF, as its location may have changed from saddle/crown. To our knowledge, there are no general expressions to estimate SCF around the brace axis accurately. As artificial neural networks (ANN) can approximate the trend of complex phenomena better than conventional data fitting, a mathematical model based on ANN is proposed to estimate SCF based on the weights and biases of trained ANN. Nine hundred thirty-seven finite element simulations were performed to generate SCF data for training the ANN. This ANN was used to model an empirical equation for SCF. The proposed empirical model can estimate SCF around the brace axis with less than 5% error. The current study provides a roadmap to using FEA and ANN for empirical modeling of SCF in tubular joints, and this approach can be applied to any joint type, with or without design modification or damage. Once a database of similar equations is available, it can be utilized for quickly estimating SCF instead of costly experimentation and FEA. Optimization of the ANN can further improve the accuracy of the developed mathematical model.

Deep drawing is characterized by very complicated deformation affected by the process parameter values including die geometry, blank holder force, material properties, and frictional conditions. The aim of this study is to model and optimize the deep drawing process for stainless steel 304 (SUS304). To achieve the purpose, die radius, punch radius, blank holder force, and frictional conditions are designated as input parameters. Thinning, as one of the major failure modes in deep drawn parts, is considered as the process output parameter. Based on the results of finite element (FE) analysis, an artificial neural network (ANN) has been developed, as a predictor, to relate important process parameters to process output characteristics. The proposed feed forward back propagation ANN is trained and tested with pairs of input/output data obtained from FE analysis. To verify the FE model, the results obtained from the FE model were compared with those of several experimental tests. Afterward, the ANN is integrated into a simulated annealing algorithm to optimize the process parameters. Optimization results indicate that by selecting the proper process parameter settings, uniform wall thickness with minimum thinning can be achieved.

The common use of pressure vessels (PVs) and pipelines and the high cost of their failure or overdesign in the oil and gas industry makes predicting their burst pressure critical. However, experimental tests of burst pressure are expensive and finite element analysis (FEA) is time consuming. Artificial neural networks (ANNs) hold the promise of rapid prediction of burst pressure for a wide range of PV materials and geometries, including pipelines with defects such as corrosion. This paper demonstrates ANNs designed and trained to accurately predict the burst pressure of both thick and thin-walled PVs for various strain hardening carbon steels. To accomplish this, we trained single and double hidden layer ANNs on experimental data, augmented with FEA, data of predicted burst pressure to create a larger database. A statistical study of hundreds of models exploring the key hyperparameters provided statistically optimal hyperparameters, resulting in highly accurate ANNs. The results, two ANNs with comparably low prediction error across geometries ranging from 3 < Do/t < 120, demonstrate the first use of ANNs to address both thick and thin-walled PV burst pressure within a single network. This is an important step towards designing ANNs for predicting the burst pressure of PVs and pipelines with defects.

Prediction of failure stages for double lap joints using finite element analysis and artificial neural networks

PurposeStress concentration factors (SCFs) are commonly used to assess the fatigue life of tubular T-joints in offshore structures. SCFs are usually estimated from parametric equations derived from experimental data and finite element analysis (FEA). However, these equations provide the SCF at the crown and saddle points of tubular T-joints only, while peak SCF might occur anywhere along the brace. Using the SCF at the crown and saddle can lead to inaccurate hotspot stress and fatigue life estimates. There are no equations available for calculating the SCF along the T-joint's brace axis under in-plane and out-of-plane bending moments.Design/methodology/approachIn this work, parametric equations for estimating SCFs are developed based on the training weights and biases of an artificial neural network (ANN), as ANNs are capable of representing complex correlations. 1,250 finite element simulations for tubular T-joints with varying dimensions subjected to in-plane bending moments and out-of-plane bending moments were conducted to obtain the corresponding SCFs for training the ANN.FindingsThe ANN was subsequently used to obtain equations to calculate the SCFs based on dimensionless parameters (α, β, γ and τ). The equations can predict the SCF around the T-joint's brace axis with an error of less than 8% and a root mean square error (RMSE) of less than 0.05.Originality/valueAccurate SCF estimation for determining the fatigue life of offshore structures reduces the risks associated with fatigue failure while ensuring their durability and dependability. The current study provides a systematic approach for calculating the stress distribution at the weld toe and SCF in T-joints using FEA and ANN, as ANNs are better at approximating complex phenomena than typical data fitting techniques. Having a database of parametric equations enables fast estimation of SCFs, as opposed to costly testing and time-consuming FEA.