These slides were presented at the NAFEMS World Congress 2025, held in Salzburg, Austria from May 19–22, 2025.

Abstract

Geometry modelling plays a crucial role for the computational representation of a given system and to obtain predictions about its performance, behaviour, etc. through simulation. A range of numerical techniques can be also coupled to enable the execution of design space exploration studies for the identification of optimal designs based on a specified number of objectives and constraints. This is typically achieved through parametric geometry models, with a set of predefined design parameters used to control the dimensions and shape of the model. Currently, BRep (Boundary Representation) is amongst the most common geometry modelling methods used in industry for engineering applications, relying on a mathematical representation of 3D shapes through a collection of basic geometry elements describing the boundaries of its volume. Nonetheless, a number of limitations are associated with such an approach, including in particular: - The level of control in manipulating geometry (i.e. to what extent the design space can be explored, allowing for the assessment of designs radically different to each other and to an original baseline). - The simulation time required for the assessment of a single design point, especially when it comes to Fine Element Analysis or Computational Fluid Dynamics simulations of complex systems. - The difficulty in re-using existing data (i.e. leveraging on information obtained from previous simulations). - The difficulty in addressing at run-time geometric and topological errors deriving from the geometry manipulation and leading to problems with accuracy, scaling, and meshing amongst others. Alternative modelling techniques have been investigated with the aim of providing more freedom and flexibility to create and manipulate the geometry of a system. Whilst some of these have proven to successfully address specific limitations associated with BRep models, further work is required to use them for simulation purposes. This paper presents a novel approach based on the use of neural networks, which offer the possibility of having a single model containing both the geometrical representation of the system under study, in conjunction with the information of its performance against design metrics of interest (e.g., stress field, flow field, displacements, temperatures, etc.). A key advantage from such approach is that it allows training the neural network model with legacy simulation data to then offer a semi-instant prediction for the performance associated with synthetically generated design candidates.