For CFD, FEM, and lookup tables.

Benefits of Local Model Network Architecture

• Distinction between rule premises (if…) and rule consequents (then…)
• Easy to interpret
• Well suited for high-dimensional input spaces
• Knowledge gained about premises and consequents can be used for DoE designs

Artificial Example
 

AiF Project in cooperation with the Institute of Fluid and Thermodynamics

(AiF: Association of Industrial Research Organizations)

• Problem: New EU regulations require a higher degree of efficiency for ventilators
  • Most existing ventilators will be banned
  • New efficient sizing and optimization strategies are required
• Computational Fluid Dynamics (CFD) models meet accuracy requirements, but:
  • Numerous evaluations are necessary for optimization
  • Computationally very expensive
  • Economically unfeasible!
• Approach to solving the problem:
  • Metamodel: A model of the CFD model
  • The CFD model generates a dataset
  • A local model network (LMN) is used

Advantages of the Local Model Network for the AiF Project Tasks
 

System Identification

• Automatic determination of model complexity saves data for training
  • Akaike’s corrected information criterion (AICc) instead of validation data
• Deterministic training procedure (no initialization required)
   

Design of Experiments

• Active learning strategy
  • The structure of LMN is used to determine future CFD simulations
• Exploitation of the distinction between linear and nonlinear effects
   

Optimization

• Local quadratic models can be used
• High probability of finding the global optimum of the LMN
  • M local models => M reasonable initialization points

Construction of a Centrifugal Fan

Computational fluid dynamics (CFD) simulations were used (~4 hours per simulation)

• Target values
  • Pressure curve and efficiency curve
• Design parameters
  • Number of blades, inner and outer diameters, various angles, etc.
     

Two Metamodeling Approaches
 

1. Direct Metamodeling

• 10 inputs: Geometric parameters and volume flow.
• 2 outputs: Efficiency and pressure.

2. Indirect Metamodeling via Characteristic Curves

• 9 inputs: Geometric parameters.
• 3 or 4 outputs: Characteristic parameters of efficiency and pressure curves.

Characteristic Curve Model 

• 3 or 4 inputs: Characteristic parameters (metamodel outputs).
• 2 outputs: Efficiency and pressure.

CFD (black line), Indirect (MIMO, blue line), Indirect (multiple MISO, green line), Direct (red line)
 

Next Chapter: 5. Nonlinear Dynamic Models Back to Overview