This presentation was made at CAASE18, The Conference on Advancing Analysis & Simulation in Engineering. CAASE18 brought together the leading visionaries, developers, and practitioners of CAE-related technologies in an open forum, to share experiences, discuss relevant trends, discover common themes, and explore future issues.

Resource Abstract

An innovative graph-based generic FEA remeshing framework is introduced in this paper, which treats an FEA unstructured mesh and high-level mesh entities as special graphs, which serves as an ideal bridge to apply AI and machine learning in the solver-mesh data flow needed in CAE simulations.
Recognizing that FEA meshing is an NP-hard problem (nondeterministic polynomial time), modern FEA unstructured mesh technology always includes heuristic algorithms which makes meshing a natural process for using data based AI and machine learning to automatically explore and achieve better mesh quality to match an FEA solver’s on-the-fly needs. This paper also introduces the next generation CAE solver-mesh concepts as following: (a) AI/machine learning is coming for the next generation of CAE, and meshing is naturally a special kind of graph which makes it a good candidate as a pioneering application in CAE; (b) meshing/remeshing has a strong trend being in the entire engineering simulation life time including preprocessing, solving, post-processing, and even geometry reconstruction; (c) a PDE is not only used to solve physical fields, but also gracefully used in the meshing process to generate/modify mesh (nodes/elements) itself.
Modern machine learning and data flow libraries (e.g., Google’s TensorFlow) tend to use graphs for AI operations and reasonings, such as linear regression, logistic regression, clustering and visual graph representation. In FEA meshing, we use graphs (its node-edge concept) as a special representation in multiple levels from small scale to large scale as noted below:
(a) Traditional atomic mesh entities (e.g., mesh element and element connectivity);
(b) High level mesh entities (called atomic regions) and entities links;
(c) A model’s multiple remesh and mesh changes in its simulation life time;
(d) A model (also allowing for multiple models) and its corresponding mesh and performance’s historic change in daily regression test of the software development release cycle;
(e) End user’s CAE models (user’s privacy considerations required) and their model variations including behavior changes.
The graphs “node-edge” is the kernel to building a machine learning network to discover and apply AI based automation in CAE meshing and simulation.
Graph and shallow/deep learning can be extensively used in the above levels/scales to enhance the remeshing robustness and mesh quality. They are componentized and organized as a flexible multi-layer neural network to adapt a “more-is-better” approach to enable data warehousing. This in turn enables a sophisticated decision-making process in an automated remeshing process.
With remeshing standing in between simulation solving and meshing, there are some pre-known knowledge and geometry/mesh patterns; however, these vary model by model because of multiple factors such as initial/final geometry patterns, FEA loads/constraints and material properties, etc. These are typical patterns to be discovered by AI and adaptive learning in an automated solver-meshing process. In the meantime, meshing is generally inexpensive compared to the solver time, leaving more room for solver-meshing to do exploration, quick learning and perform adaptive decision making. These processes can be grouped as shallow learning and deep learning depending on the size of data set and the scale in simulation life cycle.
Remeshing is an integral approach in the FEA mechanical simulation life cycle, and we have successfully applied it into the ANSYS Mechanical implicit nonlinear adaptive (NLAD) simulation and fracture crack growth as a remeshing engine to address simulation based on-the-fly large distortion, topology changes, and topology optimization with smooth geometry generation.