Agenda

Welcome

Chaired By
Eduardo Oslé Dorremochea

Presentation by NAFEMS International

Authored & Presented By
David Quinn (NAFEMS Ltd.)

Chaired By
Eduardo Oslé Dorremochea

Application of machine learning for reduce modelling in order to calibrate a 1D ventilation model of a complex road tunnel network.

Presented By
Fabian de Kluijver (JVVA)

Chaired By
Eduardo Oslé Dorremochea

Authored By
Fabian de Kluijver (JVVA)
Alberto Lopez de Arriba

The tunnels of Madrid Calle30 form a complex road network over 43 km long, with 938 fans, 85 ventilation shafts, 430 anemometers, and 284 gas detection sensors. Inaugurated in 2007, its ventilation system is being refurbished as the control equipment is reaching the end of its life cycle. During this process the operation algorithms of the ventilation system are being analysed and updated, taking into consideration the modifications that have been done in the tunnels in the past 15 years. The most suitable ventilation strategies have been defined and verified locally by means of 3D CFD simulation, but the analysis on how to achieve the appropriate ventilation conditions for each zone needs to take into account the whole infrastructure. In order to do this, a 1D CFD simulation model of the whole network has been generated to analyse the performance of the ventilation system under different operation conditions, both on emergency and normal operation. The uncertainty in some parameters of the 1D model has driven the design team to automatize its calibration against the available field data. Differential Evolution method is used for calibration, obtaining a set of parameters that minimizes an error function between the model prediction and the field data. Field data used for calibration considers different states (traffic, ambient conditions) without mechanical ventilation. However, the size and complexity of the infrastructure results in a high dimensional parameter space (202 parameters to calibrate). In this situation, if the exploration of the parameter space, during the calibration process is performed directly by using the 1D model, the total computational time would be unaffordable. To solve this issue, a surrogate model of the 1D model has been built on Python, based on the machine learning method known as Gradient Boost Decision Tree. This surrogate model has been trained with a sample of 3,000 1D simulations, where input variables (both measurable, as traffic or ambient conditions, and parameters to calibrate as tunnel roughness, traffic aerodynamic parameters, etc) are modified randomly within ranges given by expert knowledge. The surrogate model is then used to calibrate the uncertain parameters of the 1D CFD model against the available field data. With the 1D model correctly calibrated for natural ventilation, the mechanical ventilation system, defined according to the project and as-built information, is checked and calibrated against the data obtained at the field tests performed during the project. With this validation procedure it has been checked that the model is reliable as it produces results equivalent to those obtained at the field tests for all the different ventilation systems existing in the network. This calibrated 1D model will be used to evaluate current ventilation procedures and their possible improvement measures, aiming for a complete automatization of the system that ensures health and safety at the tunnel network. The methodology exposed can be used not only for calibrating ventilation models of tunnel networks but for calibrating any complex physical model with a high number of parameters. It needs to be considered that the methodology used considers a pseudo-steady state assumption being the surrogate model trained for different steady state situations. However, it is possible, by including the derivates of the variables in the surrogate model, to extend the methodology to take into account transient effects. Also, although for this study the best fit was achieved with a Gradient Boost Decision Tree regressor, it would be possible to find other regressors (as for example Neural Networks) that achieve a better behaviour for other type of systems.

How can classic engineering methods be combined with machine learning algorithms to reduce cabin noise?

Authored & Presented By
Annika Schmidt (Technical University of Applied Sciences Wildau)

Chaired By
Eduardo Oslé Dorremochea

In the quest to enhance passenger comfort by reducing cabin noise, this study investigates the integration of machine learning methods in the selection of optimal absorber/soft mount combinations in the mounting of an engine in an air taxi model. The study deals with the challenges of engine imbalances and their impact on cabin acoustics and provides a differentiated approach to conventional noise mitigation techniques. The simulation model replicates the simplified structure of an air taxi, including an integrated engine to calculate the sound pressure levels in the cabin via a fluid-structure coupling with MSC Nastran. To perform a DOE study, the structural parts of the model are reduced, and a regression model is used to derive sound pressure levels from the accelerations of the engine mounting. The developed model enables an evaluation of the noise-reducing measures based on the sound pressure level in the aircraft cabin. To model different absorber and soft mount configurations at various points of the engine mounting, Latin Hypercube Sampling was used to develop a data set containing different soft mount stiffnesses and damping, as well as absorber masses and design approaches. Several cases are considered in the data generation process to obtain the most cost-effective solution, which may include not using a noise-reducing measure at every connection point of the engine mounting. The data generated in this way was standardized and categorized in advance of the machine learning processes. Dimensionality reduction techniques such as decision trees were used to reduce the size of the data set and identify optimal combinations. The data analysis showed that the use of a cost matrix in combination with the achieved sound pressure level is extremely efficient. The data sets based on decision trees made it possible to identify target-oriented solutions. The solution selected according to the developed method led to a significant improvement in the sound pressure level, with minimal additional mass for the absorbers. In comparison, a neural network for predicting the sound pressure level was set up in combination with the Nelder-Mead-Simplex algorithm for optimizing the soft mount and absorber properties as input parameters. It has been shown that the decision tree as a transparent, comprehensible machine learning method leads to a meaningful result faster than a classical optimizer.

Questions & Answers

Chaired By
Eduardo Oslé Dorremochea

Sponsor Presentation: Dassault Systèmes

Chaired By
Eduardo Oslé Dorremochea

Analysis of Insertion and Extraction Processes.

Presented By
Javier Reboul (Principia Ingenieros Consultores S.A.)

Chaired By
Juncal Guerrero Muñoz

Authored By
Javier Reboul (Principia Ingenieros Consultores S.A.)
Javier Rodríguez

In the assembly process of many products a part is forced into another one. In general, the designer aims to minimise the insertion force, and maximise the extraction force or break a tamper seal to guarantee the desired irreversibility of the process. The simulations are typically challenging because of the associated nonlinearities, arising from frictional sliding contacts, and from the large strains and inelastic material behaviour, possibly with damage up to failure. Moreover, although the processes are essentially quasi-static (i.e., inertial forces are negligible), the residual stresses affecting the extraction scenario vary in time driven by rheological material behaviours, especially in polymers. Three case studies are described in the communication. First, the insertion of plastic and metal trim fasteners for vehicle interiors in the automotive industry. Second, the insertion and extraction processes of non-refillable closures for spirits. Third, the assembly and opening processes of a cap, tamper seal and leak seal on a biomedical screw flask. Besides, related contact features, such as interference and fluid pressure penetration to assess leakage, are also discussed. In most cases, purely static implicit simulations are possible, but explicit simulations (quasi-static with mass scaling to reduce the computational cost) are typically preferred because they provide a more robust environment for comparing different designs. The simulations help to interpret the experimental results, to identify and understand the main parameters affecting the product performance, and to optimise the design.

A cost-effective Cold Roll-Forming FE model for industrial applications.

Authored & Presented By
Timothy Senart (CRM Group)

Chaired By
Juncal Guerrero Muñoz

This paper introduces a time-effective numerical methodology for the study of steel sheet cold roll-forming processes. The output of the model and its assumptions are validated though an experimental campaign. Cold Roll Forming (CRF) is one of the most productive processes for manufacturing thin-walled products with constant cross section. It is widely used in various industrial fields with large production volumes. However, CRF processes are affected by different problems, such as plastic instabilities, twist or bow defects and elastic spring back. Numerical analyses are employed on the industrial level as an efficient and economical tool to examine the process. Their aim is to predict these defects and optimize the roll-forming line. The objective of this research is to develop a numerical model, which guarantees a good experimental correlation and a short analysis lead-time. An implicit Finite Element (FE) model is created using the commercial software LS-Dyna, to simulate the cold roll-forming process of a U-shaped profile requiring 5 forming stations. The FE model computes only a small portion of the sheet at each iteration. This is possible via particular boundary conditions which are introduced to simulate the blank continuity in the roll-forming direction. Furthermore, the use of a commercial FE software makes the methodology highly flexible and interfaceable with different problematics. The numerical model output and its assumptions are compared to experimental tests on 3.00 mm thick sheets of CP800 grade steel. Strain gauges and 3D scan are employed to study the complex strain field and kinematics of the process. Longitudinal strain, sheet geometry inside the CRF line and cross-section after springback are employed to correlate the FE model results to experimental data. The advantages and limitations of this methodology are evaluated thoroughly thanks to the experimental data. The analysis results show a good correlation within a competitive lead time for industrial applications.

Simulación CFD-DEM en el diseño de gasificadores de biomasa de lecho fluidizado.

Presented By
Pedro Gómez Molina (ICEMM SL)

Chaired By
Juncal Guerrero Muñoz

Authored By
Pedro Gómez Molina (ICEMM SL)
Álvaro José Martínez González

La fluidificación de partículas sólidas es un proceso complejo que implica la interacción entre las partículas del lecho, y entre estás y el fluido que provoca la fluidificación. La simulación numérica de este tipo de procesos se aborda mediante el acoplamiento de técnicas DEM-CFD. Aplicado al caso de gasificadores de biomasa, la simulación de este proceso permite conocer la altura del lecho y velocidad mínima de fluidificación con objeto de garantizar una correcta combustión, una mínima pérdida de carga térmica y un bajo consumo energético. El proyecto presentado muestra la optimización de un gasificador industrial para permitir la extracción automática de impropios o restos de la combustión. Para ello, se ha diseñado una parrilla de impulsión que permite la extracción de impropios, manteniendo las condiciones óptimas de fluidificación y de consumo energético por pérdidas de carga. El proyecto ha constado de 3 etapas: • Simulación CFD del distribuidor de aire: en esta etapa se analiza la pérdida de carga en el nuevo distribuidor y la uniformidad de caudales en los ramales. Tras el análisis inicial, se procede a la optimización geométrica con objeto de homogenizar el flujo y reducir la pérdida de carga. • Simulación CFD-DEM acoplada del proceso de fluidificación: esta etapa se divide en dos fases, una primera donde se realiza la calibración de los modelos de simulación con los equipos a escala, con objeto de definir una metodología contrastada y validada, y una segunda fase, donde se trabaja con el modelo de reactor a tamaño completo, tanto en la configuración inicial como en la optimizada con el nuevo diseño de distribuidor. • Simulación del sistema de extracción de impurezas del reactor: en esta etapa se realiza una simulación DEM sin fluido acoplado para validar el diseño del tornillo de extracción y sus condiciones de funcionamiento. La simulación permite obtener la tasa de extracción de las impurezas de los materiales. El proyecto se realiza con el software Open-Source MFiX para todas las simulaciones de partículas, acopladas o desacopladas con el fluido, y con un solver comercial de volúmenes finitos para la simulación CFD del sistema de distribución. Además, se realizó un cálculo estructural (cargas ambientales, térmicas y sísmicas) del reactor mediante elementos finitos empleando el software Abaqus Standard.

Body in White stiffness prediction and Optimization.

Presented By
Dimitrios Drougkas (BETA CAE Systems SA)

Chaired By
Juncal Guerrero Muñoz

Authored By
Dimitrios Drougkas (BETA CAE Systems SA)
Fabiola Cavaliere

Torsional stiffness and torsional angle are among the most important key values in a vehicle’s Body in White (BiW) development. These two values along with the first torsional stiffness mode, describe the rigidity of the vehicles body and determine its behavior concerning handling, NVH, and the basis on which the suspension components will be designed upon. Control over these values is key for further control of the vehicle’s comfort and handling. During development of a BiW, multiple modifications may occur in the early stages of design, such as part thickness changes, part geometry changes, connections positions etc. Modifications of this nature would require CAD redesign and pre-processing, in order to create a new Simulation model, and analysis re-runs, to extract and evaluate the response changes. This time-consuming process can occur multiple times during product development. In order to speed up this process, predictive models can be created to predict desired key values based on such modifications. This way the identification of important values like torsional stiffness and angle can be predicted using a trained Machine Learning model in a fraction of the time needed for the redesign and re analysis process. Such functionality enables multiple “what if” studies without the expense in analysis and design time. Well trained Machine Learning models with good accuracy can replace the Finite Element Analysis and predict the results in a fraction of the solution time. This way Optimization studies become much more approachable, since they can be done assisted by Machine Learning. As a result, we get faster optimization and more optimization studies. These methodologies have been implemented by OEM companies such as SEAT S.A., with their internally developed ARCO, the Advanced Real-time Car-body Optimization that is a physics-informed ML technique for improving car-body structure NVH behavior. ARCO is an innovative methodology to assess the noise and vibration (NVH) performance of car-body structures. Employing this methodology allows engineers to navigate the design space more efficiently, enabling early-stage decision-making and minimizing the risk of encountering design challenges in later stages. A current challenge is to scale up the methodology to real industrial models and allow an easy interaction of ARCO with standard commercial software. To that end an ongoing collaboration with BETA CAE Systems software (ANSA and KOMVOS modules) is presented, such that a user-friendly frontend interface can be used by engineers in daily activities.

Questions and Answers

Chaired By
Juncal Guerrero Muñoz

Sponsor Presentation: Beta CAE

Chaired By
Marcos Chimeno Manguán

Structural Topology Optimization as a Convolutional Neural Network.

Authored & Presented By
Raul Llamas (Airbus Defence & Space)

Chaired By
Marcos Chimeno Manguán

This presentation shows the mathematical linkage between Fully Stressed Based Structural Topology Optimization and Convolutional Neural Networks. The profound and unexpected relationships between these two subjects unveil new research opportunities and offer mutual insights into the methods employed for dimensionality reduction in domains such as "Tiny ML" and structural optimization and sizing. Moreover, applications featuring concise code written in the Julia language will demonstrate various approaches to enhance computational efficiency in addressing these problems.

The Atmospheric Boundary Layer. Discussion and methodology proposal for proper development of numerical wind tunnels.

Authored & Presented By
Francisco Hurtado (CADE soluciones de ingenieria)

Chaired By
Marcos Chimeno Manguán

The Atmospheric Boundary Layer (ABL) refers to the thin region of the atmosphere adjacent to the surface, where wind speed and turbulence are affected by the roughness and characteristics of the terrain, developing a logarithmic profile close to the surface where the velocity increases with height. The ABL is a critical factor in structural design and safety analysis for renewable energy and hydrogen sectors. On the one hand, the accurate determination of the wind loads acting on solar power plants and wind turbines is essential to ensure their safety and proper performance. On the other, the wind speed and turbulence profiles define the boundary conditions in the analysis of leakages diffusion to the environment from compressed hydrogen storages. Despite the design codes for wind loads (Eurocode 1, ASCE 7-22) include the characterization of the ABL as a function of the terrain exposure, those codes do not stablish methodologies or standards to carry out a proper numerical wind tunnel by means of CFD simulation. Therefore, the verification and validation by comparison with experimental results is recommended. As long as the design codes do not define those standards, at this presentation the challenges for the simulation and validation of the ABL are discussed, in order to define a proper methodology for the development of numerical wind tunnels and increase the confidence in the numerical results from CFD simulation. With this aim, firstly, a sensitivity analysis is carried out in order to determine the influence of the discretization schemes and computing procedures (CPU vs GPU) over the simulation of the ABL and the resulting wind loads. Secondly, a proper methodology to simulate the ABL in the numerical analysis of structures, leakages and plumes is proposed. Finally, a case study for a solar power plant is presented and the methodology proposed is validated by comparing the resulting wind load pressure coefficients from the numerical simulation with those obtained from an experimental wind tunnel.

Questions and Answers

Chaired By
Marcos Chimeno Manguán

Concluding Remarks

Chaired By
Marcos Chimeno Manguán

Thermodynamics-informed Neural Networks in Physical Simulation.

Authored & Presented By
Alberto Badias (Universidad Politécnica de Madrid)

Chaired By
Rocío Núñez

In this work we present a new method to accelerate the process of simulating the behavior of physical systems by introducing artificial neural networks informed by the laws of thermodynamics. To do so, we have created a type of neural network architecture capable of, on the one hand, learning the behavior of physical systems, and on the other, being able to guarantee a set of laws that must always be fulfilled: thermodynamics. This means combining the advantages of data science with physical hypotheses, to finally create a system capable of learning with less data and ensuring that the solutions obtained are plausible and within a safe range.

Development of a reduced FE model for failure prediction of recycled composite plies structures.

Authored & Presented By
Miguel Herraez (TecnoDigital School)

Chaired By
Rocío Núñez

In this work, we present a numerical simulation strategy to predict the mechanical failure of structures made from recycled fragments of unidirectional prepreg plies (composite oriented strand board, COSB). COSB has emerged as an alternative to conventional composites for reducing prepreg waste by reusing cut-offs and excess material in the manufacturing of secondary structures. However, the complex geometry at both the micro and mesoscale, as well as its heterogeneity, complicates its detailed analytical and numerical study. In this study, a finite element simulation model is proposed to reduce this complexity and enable the analysis of structures at larger scales. This model represents the distribution of recycled chips through: (i) a quasi-isotropic approximation, (ii) a structured mesh independent of the material's mesostructure, and (iii) a semi-analytical shear-lag model for the longitudinal strength of the chips. The results of the numerical model have been validated against experimental results, demonstrating the ability to predict the effect of chip length on the tensile strength of standard specimens. The effectiveness of this model has been proven by reducing the computational cost by 1 to 2 orders of magnitude in both the preprocessing phase and solving time compared to previously reported strategies. The presented methodology has the potential to guide the development of designs for large components and structures made with COSB.

Intelligence Morphing: Using CAD morphing and machine learning to predict CAE results

Presented By
Alberto Perez (Simulexa)

Chaired By
Rocío Núñez

Authored By
Alberto Perez (Simulexa)
Victor Jaouen & Vidya Shankar

Geometric deep learning can extract relevant features from FE or CAD models to be used as an input for several neural network architectures. These features can be used to predict the performance of designs without needing to carry out time-consuming FE analyses. This work proposes a workflow in which new designs can be created using morphing and evaluated via Convolutional Neural Networks (CNN) and other algorithms to predict the structural, crash or CFD performance. This approach uses morphing to modify the position of nodes in an FE model and automatically feed that information to a geometric deep learning algorithm to extract its relevant information. This information is used in conjunction with a Machine Learning model to provide instant feedback of the physical parameters of interest, giving an agile prediction on the performance of design changes. We also show that, thanks to Physics Informed Neural Networks (PINN), physical restrictions based on Partial Differential Equations can be added to the loss function. This allows us to reach a higher degree of correlation between Machine Learning predictions and real-world experimental data or FE results. In addition, we also tackle how PINN allows for a bigger design space exploration, giving the engineer freedom to test designs that were unaccounted for during the Machine Learning training phase. Moreover, we show how the Machine Learning model can output field, scalar, vector or time history data, effectively substituting the need for an FE solver. We extend this by utilizing the Machine Learning model to carry out optimization tasks using genetic algorithms, Kriging and Gaussian regression, amongst others, to provide the most performant design given geometrical constraints. Finally, we combine all these technologies together in a single framework to seamlessly bring design, engineering and decision-making together.

Optimización multiobjetivo mediante algoritmos genéticos.

Presented By
Nicolas Escribano (ITAINNOVA)

Chaired By
Rocío Núñez

Authored By
Nicolas Escribano (ITAINNOVA)
Gabriel Beltrán Lostal
José Manuel Bielsa Gimeno & Begoña García Arregui

La optimización topológica ha experimentado un notable incremento en su relevancia en las últimas décadas, impulsada por la necesidad de generalizar la resolución de problemas y geometrías presentes en la industria. Aunque los análisis basados en el método de los elementos finitos destacan por su precisión en la resolución, carecen de optimizadores topológicos que permitan un control personalizado del algoritmo empleado. En este trabajo se presenta la optimización multiobjetivo de un separador hiperelástico del sector de la automoción que permite la transmisión de fuerzas entre ejes. Partiendo de un espacio de diseño sometido a no linealidades geométricas, las especificaciones que tiene que cumplir el componente y un modelo de simulación inicial se optimiza tanto la geometría del separador, como la selección de material hiperelástico. Esta optimización se ha llevado a cabo mediante un flujo de trabajo personalizado en lenguaje Python que, utilizando el solver de Abaqus CAE, permite crear un bucle de optimización mediante algoritmos genéticos. Las principales ventajas de este flujo de trabajo son que se tiene un control completo de los diferentes parámetros definidos en el algoritmo genético y la creación iterativa de los modelos CAE, tanto geometría como contactos, automáticamente. Este flujo de trabajo, que integra las características y el solver de Abaqus en Python, es aplicable a diferentes casos de la industria y proporciona un optimizador topológico versátil y adaptado a las necesidades específicas de cada problema

Questions and Answers

Chaired By
Rocío Núñez