OpenMDAO doesnt want to move away from the defaults - openmdao

Question

I have some geometry that I've parameterised upstream into a mesh pre-processor which produces inputs for my external code which runs. The outputs of this is a geometry defined by its inputs, calls my external solver and then the outputs are ingested from CSV file. I've gotten it to run, but in terms of optimisation certainly hasn't minimised the value of the function. In fact it's tried values all purturbed slightly away from the starting point.
I have 5 design variables which impact the geometry and are set to go in various ranges, lets call them a,b,c,d and e. The nominal starting value of each of these variables is 0.1. Its clear from the history of the files which I saved, e.g. it has tried
a | b | c | d | e
------------------
0.100001 0.1 0.1 0.1 0.1
0.1 0.100001 0.1 0.1 0.1
0.1 0.1 0.100001 0.1 0.1
0.1 0.1 0.1 0.100001 0.1
0.1 0.1 0.1 0.1 0.100001
0.1 0.1 0.1 0.1 0.1
It somehow, then decides the original starting values were optimal, which is super weird since we can view the returned value (fout) in my case, for each one in turn. I then realised that I needed to scale my return value to something much closer to 1, I've done that I get a few more comparisons, now I get the following
Inequality constraints incompatible (Exit mode 4)
Current function value: [0.90320516]
Iterations: 3
Function evaluations: 2
Gradient evaluations: 2
Optimization FAILED.
Inequality constraints incompatible
-----------------------------------
Optimal pfc_thickness: [0.1]
Optimal conductor radius: [0.10000001]
Optimal conductor thickness: [0.1]
Optimal cut height: [0.1]
Optimal cut width: [0.3745392]
Objective value: [0.90320516]
Any kind souls out there that have tried something similar and can provide feedback/tips?
My complete code is shown below.
from openmdao.api import Problem, Group, ExternalCodeComp, IndepVarComp, ScipyOptimizeDriver
import sys
import os
from make_geometry import *
file = open("data.txt", "w")
class ExternalCode(ExternalCodeComp):
def setup(self):
self.iteration = 0
self.add_input('pfc_thickness', val = 0.0)
self.add_input('conductor_radius', val = 0.0)
self.add_input('conductor_thickness', val = 0.0)
self.add_input('conductor_origin_x', val = 0.0)
self.add_input('conductor_origin_y', val = 0.0)
self.add_input('cut_height', val = 0.0)
self.add_input('cut_width', val = 0.0)
self.add_output('fout', val=0.0)
problem_name = "mdao"
self.geometry_instance = build_geometry(problem_name)
self.input_file = 'input-thermal.i'
self.output_file = 'input-thermal_csv.csv'
self.options['external_input_files'] = [self.input_file]
self.options['external_output_files'] = [self.output_file]
#os.environ['LD_LIBRARY_PATH'] = '/usr/lib64/mpich/lib:/usr/lib64'
self.options['command'] = ['/home/adavis/opt/moose/modules/combined/combined-opt','-i','input-thermal.i']
def setup_partials(self):
# this external code does not provide derivatives, use finite difference
self.declare_partials(of='*', wrt='*', method='fd')
def __read_csv(self,filename):
import csv
with open(filename) as csv_file:
csv_reader = csv.reader(csv_file, delimiter=',')
line_count = 0
for row in csv_reader:
data = row
return data[1] # this is the maxtemp
def compute(self, inputs, outputs):
self.iteration = self.iteration + 1
# set the variables
print(inputs)
pfc_thickness = inputs['pfc_thickness'][0]
conductor_radius = inputs['conductor_radius'][0]
conductor_thickness = inputs['conductor_thickness'][0]
#conductor_origin_x = inputs['conductor_origin_x']
#conductor_origin_y = inputs['conductor_origin_y']
cut_height = inputs['cut_height'][0]
cut_width = inputs['cut_width'][0]
# set the geometry instance
self.geometry_instance.pfc_thickness = pfc_thickness
self.geometry_instance.conductor_radius = conductor_radius
self.geometry_instance.conductor_thickness = conductor_thickness
self.geometry_instance.conductor_origin_x = 5.0
self.geometry_instance.conductor_origin_y = 5.0
self.geometry_instance.cut_height = cut_height
self.geometry_instance.cut_width = cut_width
if os.path.isfile('mdao.e'):
os.remove('mdao.e')
if os.path.isfile('mdao.e'):
os.remove('input-thermal_csv.csv')
# build the file
self.geometry_instance.build_all()
# test to make sure mesh exists
if not os.path.isfile('mdao.e'):
print('meshing failed')
sys.exit(1)
filename = '%s_%s_%s_%s_%s.cub' % (pfc_thickness,conductor_radius,conductor_thickness,cut_height,cut_width)
os.system('cp mdao.cub %s' % (filename))
# the parent compute function actually runs the external code
super().compute(inputs, outputs)
# get the output
f_xy = self.__read_csv(self.output_file)
file.write('%s %s\n' % (self.iteration, f_xy))
# set the output data in f_out
outputs['fout'] = f_xy
if __name__ == "__main__":
prob = Problem()
# Create and connect inputs
prob.model.add_subsystem('p', ExternalCode())
prob.driver = ScipyOptimizeDriver()
prob.driver.options['optimizer'] = 'SLSQP'
prob.model.add_design_var('p.pfc_thickness', lower=0.1, upper = 2)
prob.model.add_design_var('p.conductor_radius', lower=0.1, upper = 3.5)
prob.model.add_design_var('p.conductor_thickness', lower=0.1, upper = 1.0)
prob.model.add_design_var('p.cut_height', lower=0.1, upper = 1.5)
prob.model.add_design_var('p.cut_width', lower=0.1, upper = 3.0)
prob.model.add_objective('p.fout')
prob.driver.options['tol'] = 1e-9
prob.driver.options['disp'] = True
prob.model.approx_totals()
# run the ExternalCodeComp Component
prob.setup()
prob.set_solver_print(level=2)
# Set input values
prob.set_val('p.pfc_thickness', 0.1)
prob.set_val('p.conductor_radius', 0.1)
prob.set_val('p.conductor_thickness', 0.1)
prob.set_val('p.cut_height', 0.1)
prob.set_val('p.cut_width', 0.1)
prob.run_driver()
print('Optimal pfc_thickness: ', prob.get_val('p.pfc_thickness'))
print('Optimal conductor radius: ', prob.get_val('p.conductor_radius'))
print('Optimal conductor thickness: ', prob.get_val('p.conductor_thickness'))
print('Optimal cut height: ', prob.get_val('p.cut_height'))
print('Optimal cut width: ', prob.get_val('p.cut_width'))
print('Objective value: ', prob.get_val('p.fout'))
file.close()

Answer 1

Is your output file ('input-thermal_csv.csv') being properly regenerated on each execution? I ask because it looks like you intend to delete it in your compute, but the condition seems incorrect so it would never get deleted:
if os.path.isfile('mdao.e'):
os.remove('mdao.e')
if os.path.isfile('mdao.e'):
os.remove('input-thermal_csv.csv')

Answer 2

When you use finite difference, you need to be very careful about step sizes and methods. You can consider switching to relative instead of absolute step size, or using a different step size (default in OpenMDAO V3.9 is 1e-6, absolute). If you are using file i/o then be sure you are writing output with sufficient number of digits as well. If you are truncating the output to say 8 or 9 digits, then you might be getting effectively 0 change for very small steps (because the change is down in the digits you are not printing to the file).
You can have a look at the options in the declare_partials and check_partials methods to see how to adjust these settings.