Capstone design(1)

1. Project goal

Optimize vent hole(outlet) pattern for reducing the cooling drag of UAV's motor efficiency.

2. Project structure overview

2.1 Surrogate modeling

1. Extract arbitrary outlet data from CFD simulation(ANSYS Fluent)
2. Upscale the input data for high-dimensional surrogate modeling
3. PCA(Principal Component Analysis) for dimension reduction(for train quality)
4. Train the surrogate model with the input data and output data

• Input data: Vent-hole matrix
• Output data1: Drag
• Output data2: Motor surface temperature(avg)
• Output data3: Motor surface temperature(max)

Note: Drag(Cd) and surface temperature(Tavg, mean) are conflicting objectives. So, we decided to optimize this multi-objective problem using Genetic-algorithm.

2.2 Generate vent hole pattern sets

1. Pattern type selection - defined by transformation group
   • Grid
   • Circular
   • Corn
2. Shape definition - define by closed area function(user-defined) F(x, y) <= 0
3. Chromosome(pattern) generation - generate vent hole pattern sets
4. Generate pattern matrix - Model's input data Example of generated patterns:

2.3 Genetic algorithm

1. Initialize population - can be customized
   • Shape gene distribution
   • Pattern gene distribution
   • population size, ...etc
2. Evaluate fitness - can be customized
   • Fitness1: Drag
   • Fitness2: Motor surface temperature(avg)
   • Fitness3: Motor surface temperature(max)
3. Selection - can be customized
   • Tournament selection
   • Roulette wheel selection
   • Stochastic universal sampling
   • Elitism
4. Crossover - can be customized
   • Single-point crossover
   • Two-point crossover
   • Uniform crossover

5. Mutation - can be customized
   • Gaussian mutation(stochastic approach)
   • Random mutation(probabilistic approach)
6. Termination - can be customized
   • Max generation
   • Population diversity
   • Extraordinary fitness value
   • ...etc

3. Project running

1. Clone this repository

   git clone [repo link]

2. Install virtual environment

   vscode based(recommended):

   • ctrl + shift + p -> Python: Create environment.
   • select ./venv directory.
   • select python interpreter version, 3.11.0 or higher recommended.

3. Install required packages

   pip3 install -r packages.txt

4. Run the main script

   1. goto experimental_template.ipynb file(make sure you have installed jupyter notebook)

   2. adjust genetic algorithm hyperparameters

   3. run the script

      suite = GAPipeline[VentHole](
          suite_name="exp/tournament/config", 
          # 🔼 Name of experiment, and also the directory name store/{suite_name}
      
          suite_max_count=50, 
          # 🔼 Maximum generation
      
          suite_min_population=20, 
          # 🔼 Minimum population size
      
          suite_min_chromosome=40, 
          # 🔼 Minimum unique chromosome size
      
          crossover_behavior=UniformCrossover(), 
          # 🔼 Crossover behavior
      
          selector_behavior=TournamentSelectionFilter(tournament_size=4), 
          # 🔼 Selection behavior
      
          fitness_calculator=VentFitnessCalculator( 
              # 🔼 Fitness model
              model_trainer_tuple=(
                  gpr_model_trainer,
                  gpr_model_trainer,
                  gpr_model_trainer,
              ),
              criteria_weight_list=CRITERIA_WEIGHT,
              drag_criterion=DRAG_CRITERION,
              drag_std_criterion=DRAG_STD_CRITERION,
              avg_temp_criterion=AVG_TEMP_CRITERION,
              avg_temp_std_criterion=AVG_TEMP_STD_CRITERION,
              max_temp_criterion=MAX_TEMP_CRITERION,
              max_temp_std_criterion=MAX_TEMP_STD_CRITERION,
          ),
          immediate_exit_condition=lambda x: x[0] >= 0.725 and x[1] >= 0.725, 
          # 🔼 Exit condition, extra ordinary fitness value
      
          mutation_probability=0.01,  # 1% 
          # 🔼 Mutation probability
      
          population_initializer=VentInitializer( 
              population_size=150, 
              # 🔼 Initial population size
      
              grid_scale=GRID_SCALE,
              grid_resolution=GRID_RESOLUTION,
              pattern_bound=GRID_BOUND,
      
              pattern_gene_pool=[ 
                  # 🔼 Pattern gene pool, you can adjust the gene pool
                  circular_params,
                  corn_params,
                  grid_params,
              ],
      
              shape_gene_pool=[ 
                  # 🔼 Shape gene pool, you can adjust the gene pool
                  circle_params,
              ],
          ),
      )

   4. Check the results

      Result example:

   • Gene, fitness data: check store/{suite_name} directory

     Note:

     All the results are recorded at store directory by json format data. You can visualize or analyze the results.