NAFEMS Americas and Digital Engineering (DE) teamed up (once again) to present CAASE, the (now Virtual) Conference on Advancing Analysis & Simulation in Engineering, on June 16-18, 2020!

CAASE20 brought together the leading visionaries, developers, and practitioners of CAE-related technologies in an open forum, unlike any other, to share experiences, discuss relevant trends, discover common themes, and explore future issues, including:
-What is the future for engineering analysis and simulation?
-Where will it lead us in the next decade?
-How can designers and engineers realize its full potential?
What are the business, technological, and human enablers that will take past successful developments to new levels in the next ten years?

Resource Abstract

For decades, the role of numerical simulation in product development has been growing dramatically. Today, even more prominent with the rise of industry 4.0, simulation has shifted from being a validation tool of mature designs into a means of exploration of product design space. The growth of high performance computing infrastructure and the progress in high-fidelity simulation methods have certainly contributed to numerical simulation being key in the reduction of physical testing and in product performance improvement. Yet, the time required to run a simulation is, most of the time, a bottleneck in the engineer’s optimisation loop and for larger design spaces it can result in automated shape optimization being simply intractable. This needs to be addressed on the way to better simulation-driven design.

Unsurprisingly and like in many other disciplines, machine learning has got an answer. The promise is to leverage the untapped value of historical simulations to learn a surrogate model as a cheaper data-driven substitute for the numerical simulator. Surrogate models are used in CFD simulations as well as many other simulations such as structural analyses involving a finite element solver. Most of the existing surrogate modeling approaches rely on Gaussian Process regressors (Kriging) and are thus limited to predicting the performance of shapes with a fixed low-dimensional parameterization. On top of that, kriging methods are meant for predicting global scalar values but they are not capable of predicting fields (e.g. velocity or pressure values at every point of the shape). More recently [1], authors proposed using a deep neural network as a regressor, which takes polycube-mapped meshes as inputs. Unfortunately, this means that the method is only applicable to surface meshes that have been remapped to the same polycube topology and are composed of a single connected component. One way to overcome this issue is to rely on geometric deep learning techniques [2, 3] that directly operate on 3D meshes.

In this abstract, we give an account of the successful application of 3D geometric deep learning techniques powered by the Neural Concept Shape software to accelerate and automate shape design optimization in real industrial use cases. More particularly, we demonstrate the case of a fixed-wing drone shape optimisation. We train a deep Geometric Convolutional Neural Network (GCNN) to reproduce the flow and pressure field, normally obtained using a CFD solver, directly as a Neural Network prediction. Importantly, the input of our Neural Network can be a mesh representation of the shape since we rely on spatial mesh convolutions as described in [3]. In our software, we use geodesic coordinates for convolutions on the surface of the objects, and Euclidean coordinates for convolutions in the bulk of the domain. As opposed to existing methods, we do not use an interpolation on a regular grid or an image, and do not require prior re-meshing of the shape.

The reported approach is beneficial on many critical aspects: first, it allows us to compute approximate solutions orders of magnitude faster than the typical numerical simulators (tens of milliseconds instead of multiple hours); and, second, it allows to compute the gradients of the objective (e.g. aerodynamic efficiency) with respect to the input shape. This ultimately allows us to run optimization loops in reasonable time and makes it possible to use first-order optimization methods.