A machine learning assisted preliminary design methodology for repetitive design features in complex structures

From Design Computation
Revision as of 08:05, 12 April 2023 by Giuliapattini (talk | contribs)

Jump to: navigation, search
DCIO2022-Logo.png
DC I/O 2022 Paper by Omar A.I. Azeem and Lorenzo Iannucci. https://doi.org/10.47330/DCIO.2022.CAXL3310 | Left | Left


DCIO2022 S1 2 O-Azeem.png


Abstract

The current industrial practice used at the preliminary design stage of complex structures involves the use of multifidelity submodelling simulations to predict failure behaviour around geometric and structural design features of interest, such as bolts, fillets, and ply drops. A simplified global model without the design features is first run and the resulting displacement fields are transferred to multiple local models containing the design features of interest. The creation of these high-fidelity local feature models is highly expert dependent, and their subsequent simulation is highly time-consuming. These issues compound as these design features are typically repetitive in complex structures. This leads to long design and development cycles. Application of machine learning to this framework has the potential to capture a structural designer’s modelling knowledge and quickly suggest improved design feature parameters, thereby addressing the current challenges. In this work, we provide a proof of concept for a machine learning assisted preliminary design workflow, see Figure 1, whereby feature-specific surrogate models may be trained offline and used for faster and simpler design iterations. The key challenge is to maximise the prediction accuracy of failure metrics whilst managing the high dimensions required to represent design feature simulation parameters in a minimum training dataset size. These challenges are addressed using: a modified Latin Hypercube Sampling scheme adjusted to improve design of experiment in composite materials; a bi-linear work-equivalent homogenisation scheme to reduce the number of nodal degrees of freedom; a non-local volume-averaged stress-based approach to reduce the number of target features; and linear superposition of stacked bi-directional LSTM neural network models. This methodology is demonstrated in a case study of predicting the stresses of open hole composite laminates in an aerospace C-spar structure. Results highlight the high accuracy (>90%) and time saving benefit (>15x) of this new approach. This methodology may be used to faster correct and iterate the preliminary design of any large or complex structure where there are repetitive localised design features that may contribute to failure, such as in Formula 1 or wind turbines. Combined with exascale computing this methodology may also be applied for predictive virtual testing of digital twins.

Keywords

Machine learning, modelling, simulation, preliminary design

Bibliography

  • OSTERGAARD, M.G., IBBOTSON, A.R., ROUX, O.L., AND PRIOR, A.M. 2011. Virtual testing of aircraft structures. CEAS Aeronautical Journal 247, 1, 83.
  • BOEING 787. Article Title. Boeing 787 By Design Advanced Composite Use. https://www.boeing.com/commercial/787/by-design/#/advanced-composite-use
  • AIRBUS A350. Article Title. Airbus A350 Family. https://www.airbus.com/en/products-services/commercial-aircraft/passenger-aircraft/a350-family
  • PINHO, S.T., DARVIZEH R., ROBINSON P., SCHUECKER C. AND CAMANHO PP. 2012. Material and structural response of polymer-matrix fibre-reinforced composites. Journal of Composite Materials, 46(19-20)
  • SIMULIA. 2020. Article Title. Damage initiation for fibre-reinforced composites. https://help.3ds.com/2020/English/DSSIMULIA_Established/SIMACAEMATRefMap/simamat-c-
  • QIN, X., ET AL. 2021. Effects of countersunk hole geometry errors on the fatigue performance of CFRP bolted joints . Proc IMechE Part B: J Engineering Manufacture.
  • JOSEPH, V. R., GUL, E., AND BA, S. 2015. Maximum Projection Designs for Computer Experiments. Biometrika, 102, 371-380.damageinitfibercomposite.htm?contextscope=all&id=9291b1119c3345eba2858e79b982381f
  • FAROKHI, H., BACARREZA, O., FERRI ALIABADI, M.H., 2020. Probabilistic optimisation of mono-stringer composite stiffened panels in post-buckling regime. Structural and Multidisciplinary Optimization. 62:1395–1417
  • LIU, R., ET AL. 2017. Context Aware Machine Learning Approaches for Modeling Elastic Localization in Three-Dimensional Composite Microstructures. Integr Mater Manuf Innov 6:160–171
  • LOGARZO, H.J., CAPUANO, G., RIMOLI, J.J. 2020. Smart constitutive laws: Inelastic homogenization through machine learning. Computer Methods in Applied Mechanics and Engineering 373 113482
  • ZOU, X., ET AL. 2021. An Abaqus plugin for efficient damage initiation hotspot identification in large-scale composite structures with repeated features. Advances in Engineering Software 153 102964
  • WANG, Z., WAN, Z., GROH, R.M.J., AND WANG, X. 2021. Aeroelastic and local buckling optimisation of a variable-angle-tow composite wing-box structure. Composite Structures 258 11320
  • NIU MC-Y. 1992 Composite Airframe Structures: Practical Design Information and Data. Hong Kong: Conmilit.
  • GERON, A. (2019). Hands-on machine learning with Scikit-Learn, Keras and TensorFlow: concepts, tools, and techniques to build intelligent systems (2nd ed.). O’Reilly.