Automated FEACFD Hexahedral Mesh Generation Using an Integrated Neural NetworkRuleBased Method

1
Automated FEA/CFD Hexahedral Mesh Generation
Using an Integrated Neural Network/Rule-Based
Method
( Kickoff Meeting )
Presented At
Parker Hannifin Corporation
October 5 By Samuel H. Huang, Director Intelligent
CAM Systems Laboratory
2
Outline of the Project

• To develop a mechanism for automated Hex-mesh
  generation
• for complicated geometry by computerizing human
  expertise
• using an integrated neural network/rule-based
  approach.
• Hex-Mesh Generation Use of hexahedral meshes
  will result in
• higher analysis accuracy while requiring
  less computations.
• Need to Automate Lots of human Interaction
  required for block
• generation in available software tools.
• Neural Network Approach Extract knowledge from
  human expertise.
• Mechanism The knowledge extracted will be
  documented,
• verified, and then computerized to develop a
  prototype software tool
• that can be interfaced with existing
  multi-block mesh generation
• software tools.

3
About Neural Network

• Inspired by biological nervous systems.
• Neural network is trained to perform a particular
  function by
• learning from examples.
• Neural Networks is applied in areas such as
  pattern recognition, classification, speech
  recognition, image analysis, etc.
• The knowledge is represented at a sub-symbolic
  level, which needs to be decoded so that it will
  be intelligible to users.

Neural Network including connections (called
weights) between neurons Input Output Compare
Compare
Target
Output
Adjusted weights
4
Neuron Model

• Consists of a processing element with synaptic
  input connections and a single output.
• The neuron output signal is given by
• o f ( w t x ) f ( S wx xi ) for i 1 to
  n
• where w is the weight vector defined as
• w w 1 w2 wn
  
• and x is the input vector
• x x1 x2 xn
• Typical Activation function
• f (net) 1, net gt 0 Or f
  (net) 1, net gt 0
• -1, net lt 0
  0, net lt 0
• (Bipolar Function)
  (Unipolar Function)

5
Neural Network Application in Mesh Generation

• Research Papers
• An automatic mesh generator for handling small
  features in open boundary power transmission line
  problems using artificial neural networks, by
  Triantafyllidis and Labridis
• Parallel training of neural networks for finite
  element mesh decomposition by Topping, Khan, and
  Bahreininejad
• Parallel quadrilateral subdomain generation for
  finite element analysis, by Sziveri, Cheng,
  Bahreininejad, Cai, Thierauf, and Topping
• Finite-element mesh generation using
  self-organizing neural networks, by Manevitz,
  Yousef, and Givoli

6
Neural Network Application in Mesh Generation
(cont.)

• Research has been done in the past in mesh
• generation using neural networks.
• Most of the work is done in local problems
  such as
• predicting the number of mesh elements.
• No current research in block generation using
  neural
• networks.

7
Issues for Discussion

• Data
• - Input Data (for example, load
  considerations)
• - Data Collection
• - Data Processing
• Integration with Existing Software Tool
• - Input format (Part geometry)
• - Output format (Block topology)
• Project Planning