OpenMDAO Scripting Interface

OpenMDAO provides a programmatic interface that allows you to write a Python script that describes the structure of the model and provides the ability to interact with objects in the framework.

The goal of this section of the User Guide is to explain and demonstrate several aspects of the OpenMDAO script interface. This section is intended primarily as a reference. If you are an inexperienced user, you would best be served by reading and understanding the examples in Getting Started: A Simple Tutorial Problem and A More Complex Tutorial Problem.

OpenMDAO Fundamentals

The following sections discuss the the basics of OpenMDAO.

OpenMDAO Package Hierarchy

Package is a Python concept that provides a structure for organizing variables and functions in a logical hierarchical fashion. Packages allow you to import needed functions and class definitions into the Python environment using dotted module names, where the branch levels are separated by a period (”.”).

The OpenMDAO package hierarchy includes several subpackages, all of which are prefixed by openmdao.:

  • openmdao.main – core infrastructure for the framework
  • openmdao.lib – OpenMDAO’s standard library, containing some important plugins (drivers, traits, etc.) that are available to users of the framework
  • openmdao.units – unit definition and conversion
  • openmdao.examples – tutorials and example problems for learning OpenMDAO
  • openmdao.util – utilities used by OpenMDAO but can also be used standalone
  • openmdao.test – functions and classes used strictly for unit testing

OpenMDAO users and component developers will likely need only the first three of these (main, lib, and units). Importing classes and functions from OpenMDAO’s libraries is performed with the same syntax as loading any other Python module:

from openmdao.main.api import Component, Assembly
from openmdao.lib.drivers.api import CONMINdriver

Here, the fundamental OpenMDAO component classes Component and Assembly are loaded from openmdao.main.api, along with the CONMIN driver from openmdao.lib.drivers.api.

To simplify the imports, a selection of the most commonly used imports was placed in the pseudo-package openmdao.main.api. You can obtain a complete listing of what is available in this module by using the dir() command in Python. Likewise, a pseudo-package was also created to house some of the most commonly used imports from the standard library. In general, it contains variables and drivers. Most of these items are also explained elsewhere in the User Guide.

Importing more objects into the namespace of your module increases the likelihood of name collision, so you should import only the objects that you need. You should avoid using from <modname> import * because it puts every object from the given module into the current namespace.

from openmdao.main.api import *

import openmdao.main.api

from openmdao.main.api import Component, Assembly, Driver

Naming Conventions

Components and variables that are instantiated into the OpenMDAO model hierarchy must follow the same naming syntax as attributes in the Python language. To summarize, they can include only alphanumeric characters and the underscore, and the lead character cannot be a number. Any attempt to create a component or a variable that does not conform to Python’s syntax should result in an exception. This restriction was required because these entities essentially exist as Python variables. One unfortunate side effect is that names with spaces are not allowed. OpenMDAO checks for compliance when a variable or Component instance is created:

>>> from openmdao.main.api import Assembly
>>> from openmdao.examples.enginedesign.chassis import Chassis
>>> top = Assembly('top')
>>> top.add('chassis1',Chassis())
<openmdao.examples.enginedesign.chassis.Chassis object at ...
>>> top.add('the chassis',Chassis())
Traceback (most recent call last):
NameError: name 'the chassis' contains illegal characters

In the OpenMDAO source and examples, we’ve tried to follow the PEP 8 standard, which specifies a naming convention for component instance names and variable names. For all variable names, PEP 8 prescribes the use of lower case names with words separated by underscores. Naturally, PEP 8 compliance is not a requirement that will be forced on users, but it is a good style guideline.

Creating New Components

Components are the basic building block of the OpenMDAO model, so you need to be familiar with how to create and execute them. The concept of the component and the place it holds in the OpenMDAO architecture is given in the Introduction to the OpenMDAO Framework.

Presumably you have your own components to implement in OpenMDAO as part of a larger model or process. This implementation will usually require the creation of an OpenMDAO Python component based on the Component class and conforming to the Component API.

The Component API

Every component in the OpenMDAO framework is an object that conforms to a specific interface. At present, the easiest way to match this interface is to inherit from the built-in Component class and then override the execute() function to give the component some kind of run behavior. Likewise, the __init__() function can also be overridden to prescribe the component’s behavior when it is instantiated. This is mostly useful for defining any internal private variables that need to be saved between runs but aren’t needed by other components in the framework.

A simple component that implements an equation with two inputs is shown below:

from openmdao.main.api import Component
from openmdao.lib.datatypes.api import Float

class Equation(Component):
    """ Evaluates the equation f(x,y) = (x-3)^2 + xy + (y+4)^2 - 3 """

    # Component Input
    x = Float(0.0, iotype='in', desc='The variable y')
    y = Float(0.0, iotype='in', desc='The variable x')

    # Component Output
    f_xy = Float(0.0, iotype='out', desc='F(x,y)')

    # Initialization function (technically not needed here)
    def __init__(self):
        super(Equation, self).__init__()

    # Executes when component is run
    def execute(self):
        """ Solve (x-3)^2 + xy + (y+4)^2 = 3
        Optimal solution (minimum): x = 6.6667; y = -7.3333

        x = self.x
        y = self.y

        self.f_xy = (x-3.0)**2 + x*y + (y+4.0)**2 - 3.0

In this example, the __init__() function doesn’t do anything but call the equivalent in the base class, so technically it should be removed from this class definition.

One additional function that may need to be defined in certain cases is save_to_egg(). Sometimes a wrapped code might require some additional files or directories to be packed with it. These kinds of things can be taken care of in save_to_egg(). It is important not to forget to call the save_to_egg() for the base class.


save_to_egg example


In OpenMDAO, a variable is an attribute that can be seen or manipulated by other entities in the framework. Any data that is passed between components in a model must use variables to declare the inputs and outputs for each component.

You can create a variable for a component in two ways. The first is to declare it in the component’s class definition as shown in the example given in the simple tutorial problem. A simple component that takes a floating point number as an input and provides a floating point number as an output would look like this:

from openmdao.main.api import Component
from openmdao.lib.datatypes.api import Float

class Simple(Component):
    """ A simple multiplication """

    # set up interface to the framework
    x = Float(1.0, iotype='in', desc='The input x')
    y = Float(iotype='out', desc='The output y')

    def execute(self):
        """ y = 3*x """

        self.y = 3.0*self.x

The example above shows the way the majority of users will create variables. An alternative way to declare them is to use the add_trait function that is part of the Component public interface. First, lets define the same class in the shell but without the variables x and y.

from openmdao.main.api import Component
from openmdao.lib.datatypes.api import Float
class Simple(Component):
    """ A simple multiplication """
    def execute(self):
        """ y = 3*x """
        self.y = 3.0*self.x

Next, the add_trait function is used to add the input x and the output y after an instance of Simple has been created:

>>> equation = Simple()
>>> equation.add_trait('x',Float(1.0, iotype='in', desc='The input x'))
>>> equation.add_trait('y',Float(iotype='out', desc='The output y'))
>>> equation.x=7
>>> equation.y

The primary use of add_trait is to create a variable dynamically at some point after the component has been created (possibly during execution).

>>> from openmdao.examples.simple.paraboloid import Paraboloid
>>> from openmdao.lib.datatypes.api import Int
>>> test=Paraboloid()
>>> test.z
Traceback (most recent call last):
AttributeError: 'Paraboloid' object has no attribute 'z
>>> test.add_trait('z',Int(7777, iotype='out', desc='An Int'))
>>> test.z

Some specialized components will make use of the ability to create variables on the fly, but most general components won’t need this.

The example above shows how to directly access a variable, but there is also an indirect access using a set and get method. Set and get are primarily used by the framework to pass data between variables. In some cases a model developer may need to use them – but only for specific cases where some objects are executing on remote servers.

Here is an example of the get function:

>>> from openmdao.examples.enginedesign.engine import Engine
>>> my_engine = Engine()
>>> my_engine.bore
>>> my_engine.get("bore")

Here is an example of the set function:

>>> my_engine.RPM = 2500
>>> my_engine.RPM
>>> my_engine.set("RPM",3333)
>>> my_engine.RPM


The underlying implementation of variables in OpenMDAO was accomplished through a Python add-on called Traits. Traits provide a way to apply explicit typing to the normally untyped Python attributes. They also provide the capability to add some other features to the variables, including unit checking and conversion, default values, upper and lower bounds, and a way to create callback functions that execute under specified conditions.

In general, you won’t need to worry about traits or how variables are implemented, but those of you who want to create custom datatypes can do so by defining a new custom trait. More details on traits can be found on Enthought’s Traits project page.

Built-in Variable Types

Summary of Variable Types

Nam Callable Signature
Arr Array( [default_value = None, shape = None, value = None, dtype = None, units = None, iotype = None, desc = None, units = None] )
Bool Bool( [value = None, desc = None, iotype = None] )
Complex Complex( [value = None, desc = None, iotype = None] )
Enum Enum( [default_value, values = (), desc = None, iotype = None, aliases = ()] )
File File( [default_value = None, iotype = None, desc = None, low = None, high = None, path = None, content_type = None, binary = False, local_path = None] )
Float Float( [default_value = None, iotype = None, desc = None, low = None, high = None, exclude_low = False, exclude_high = False, units = None] )
Instance Instance( [klass = None, desc = None, iotype = None, factory = None, args = None, kw = None, allow_none = True, adapt = None, module = None, required = False] )
Int Int( [default_value = None, iotype = None, desc = None, low = None, high = None, exclude_low = False, exclude_high = False] )
Range Deprecated. Use OpenMDAO’s Int or Float.
Str Str( [value = None, desc = None, iotype = None] )

A more detailed list of Enthought’s Traits is given in their documentation. Traits are also available for use as variables in the framework, though we haven’t included examples of the more exotic ones. If you need to use one, remember that iotype and desc should be added to the arguments when one of these is instantiated. The traits use **metadata to store these user-defined attributes.

A variable is declared with a number of arguments, many of which are optional.

The iotype attribute is required for all variables regardless of type. Its sole function is to tell the framework whether the variable should be treated as an input or an output. Presently, the only two options for this attribute are 'in' and 'out'.

Summary of iotypes

iotype Description
iotype='in' Component input
iotype='out' Component output

The desc attribute is a concise description of the variable – one or two sentences should be fine. While nothing in the framework requires this description, it would be wise to include one for every input and output of your components.

It is possible to create new types of variables to use in your models. For an example of a user-created variable, see Building a Variable Plugin from a Python Module.


It is possible to use an array as a variable through use of the Array trait. The value for an Array can be expressed as either a Python array or a NumPy array. NumPy arrays are very useful because of NumPy’s built-in mathematical capabilities. Either array can be n-dimensional and of potentially any type.

Constructing an Array variable requires a couple of additional parameters that are illustrated in the following example:

>>> from openmdao.lib.datatypes.api import Array
>>> from numpy import array
>>> from numpy import float as numpy_float
>>> z = Array(array([[1.0,2.0],[3.0,5.0]]), dtype=numpy_float, shape=(2,2), iotype='in')
>>> z.default_value
array([[ 1.,  2.],
       [ 3.,  5.]])
>>> z.default_value[0][1]

Here, we import the Array variable and the NumPy array, which is a general-purpose n-dimensional array class. A 2-dimensional array is assigned as the default value for the variable named z.

The dtype parameter defines the type of variable that is in the array. For example, using a string (str) for a dtype would give an array of strings. Any of Python’s standard types and NumPy’s additional types should be valid for the dtype parameter. The alternate typecode specification is also supported for non-NumPy arrays (e.g., typecode='I' for unsigned integers.)

The shape parameter is not a required attribute; the Array will default to the dimensions of the array that are given as the value. However, it is often useful to specify the size explicitly, so an exception is generated if an array of a different size or shape is passed into it. If the size of an array is not determined until runtime (e.g., a driver that takes an array of constraint equations as an input), then the shape should be left blank.

An array can also have a single unit defined with the units parameter. This unit applies to every element in the array, and it enables unit checking and conversion when connecting an array output to an array input.

Below is an example of a simple component that takes two Arrays as inputs and calculates their dot product as an output.

from numpy import array, sum, float

from openmdao.main.api import Component
from openmdao.lib.datatypes.api import Array, Float

class Dot(Component):
    """ A component that outputs a dot product of two arrays"""

    # set up interface to the framework
    x1 = Array(array([1.0,2.0]), dtype=float, desc = "Input 1",
    x2 = Array(array([7.0,8.0]), dtype=float, desc = "Input 2",

    y = Float(0.0, iotype='out', desc = "Dot Product")

    def execute(self):
        """ calculate dot product """

        if len(self.x1) != len(self.x2):
            self.raise_exception('Input vectors must be of equal length',

        # Note: array multiplication is element by element
        self.y = sum(self.x1*self.x2)

        # print the first element of x1
        print x1[0]

Multiplication of a NumPy array is element by element, so sum is used to complete the calculation of the dot product. Individual elements of the array can also be accessed using brackets. An OpenMDAO Array behaves like a NumPy array, so it can be used as an argument in a NumPy function like sum.

Note that this is a horrible way to do a dot product. Numpy has a dot function which is much faster than sum.


It is possible to use an Enum (enumeration) type as a variable in OpenMDAO. This is useful for cases where an input has certain fixed values that are possible. For example, consider a variable that can be one of three colors:

from openmdao.lib.datatypes.api import Enum
from openmdao.main.api import Component

class TrafficLight(Component):
    color2 = Enum('Red', ('Red', 'Yellow', 'Green'), iotype='in')

Then we can interact like this:

>>> test = TrafficLight()
>>> test.color2
>>> test.color2="Purple"
Traceback (most recent call last):
TraitError: : Trait 'color2' must be in ('Red', 'Yellow', 'Green'), but a value of Purple <type 'str'> was specified.
>>> test.color2="Green"
>>> test.color2

However, if the Enum is being used to select the input for an old code, then you will most likely need to feed it integers, not strings. To make this more convenient, the Enum includes an optional parameter alias that can be used to provide descriptive strings to go along with the numbers the code expects.

from openmdao.lib.datatypes.api import Enum
from openmdao.main.api import Component

class TrafficLight(Component):
    color = Enum(0, (0, 1, 2), iotype='in', aliases=("Red", "Yellow", "Green"))

Let’s create an instance of this component and try setting the Enum.

>>> test = TrafficLight()
>>> test.color=2
>>> test.color

If we set to an invalid value, an exception is raised.

>>> test.color=4
Traceback (most recent call last):
TraitError: : Trait 'color' must be in (0, 1, 2), but a value of 4 <type 'int'> was specified.`

We can also access the list of indices and the list of aliases directly from the trait.

>>> color_trait = test.get_trait('color')
>>> color_trait.aliases
('Red', 'Yellow', 'Green')
>>> color_trait.values
(0, 1, 2)
>>> color_trait.aliases[test.color]

If the default value is not given, then the first value of the list is taken as the default.

color2 = Enum(('Red', 'Yellow', 'Green'), iotype='in')

This is the simplest form of the Enum constructor.

It is also possible to produce a simple array that behaves like an Enum, where each element of the array can contain only a value that is in the Enum. This kind of variable can be defined by creating a List of Enums.

from openmdao.lib.datatypes.api import Enum, List
from openmdao.main.api import Component

class Dice(Component):
    roll = List( Enum(1, (1, 2, 3, 4, 5, 6)), iotype='in')

This example defines a variable named roll that can contain the values for any number of dice. Instead of giving a List as the default value, we’ve given it the definition for an Enum variable that has a default value of 1 and a set of valid values spanning the integers 1 to 6. Note that the Enum doesn’t need an iotype, but the List does.

>>> my_dice = Dice()
>>> # Valid
>>> my_dice.roll = [1, 6, 3, 2, 2]
>>> # Invalid
>>> my_dice.roll = [1, 6, 3, 2, 7]
Traceback (most recent call last):
TraitError: : Trait 'roll' must be in (1, 2, 3, 4, 5, 6), but a value of 7 <type 'int'> was specified.

File Variables

The File variable contains a reference to an input or output file on disk. It is more than just a text string that contains a path and filename; it is a FileReference that can be passed into other functions expecting such an object. FileReferences have methods for copying the reference and opening the referenced file for reading. The available “flags” are defined by FileMetadata, which supports arbitrary user metadata.

from openmdao.lib.datatypes.api import File

text_file = File(path='source.txt', iotype='out', content_type='txt')
binary_file = File(path='source.bin', iotype='out', binary=True,
                        extra_stuff='Hello world!')

The path must be a descendant of the parent component’s path, as explained in Interacting with Files and Directories. The binary flag can be used to mark a file as binary.


Provide some examples to demonstrate the options.

Instance Traits

An Instance is a trait that requires any value assigned to it to be either an instance of a specific class or an implementation of a specific Interface. The class or Interface to be matched is the first argument to the constructor. Failure to match the specified class or Interface will result in an exception being raised. Instance traits are typically used to implement Sockets, which are placeholders for plugins within a component, but they may also be used to implement Variables by setting their iotype metadata attribute to 'in' or 'out'. In this case, it is important to also set the copy metadata attribute so the framework knows how to copy the data to connected components. Allowable values for copy are 'deep' (the default), 'shallow', and None. A copy value of None indicates that the data will be passed by reference and no copy will be made.

from openmdao.main.api import Component
from openmdao.lib.datatypes.api import Instance
from openmdao.main.interfaces import ICaseRecorder, ICaseIterator

class Fred(Component):
    """ A component that takes a class as an input """

    recorder = Instance(ICaseRecorder, desc='Something to append() to.',
    caseiter = Instance(ICaseIterator, desc='set of cases to run.',

In this example, we have one Socket and one input that are Instances. The input called caseiter requires data objects that implement the ICaseIterator interface. The Socket called recorder is required to implement the ICaseRecorder interface.

The attribute required is used to indicate whether the object that plugs into a Socket is required. If required is True, then an exception will be raised if the object is not present.

Unit Conversions with Float and Array

OpenMDAO also supports variables with explicitly defined units using the Float and Array variable types, which are included as part of the Standard Library. Both types provide the following useful effects when utilized in the framework.

  • Automatically convert a value passed from an output to an input with compatible units (e.g., 'inch' and 'm')
  • Raise an exception when attempting to pass a value from an output to an input having incompatible units (e.g., 'kg' and 'm')
  • Allow values to be passed between unitless variables and variables with units; no unit conversion occurs

A complete list of the available units is given in the Appendix: Summary of Units. The unit conversion code and the base set of units come from the PhysicalQuantities package found in Scientific Python. It was necessary to add a few units to the existing ones in PhysicalQuantities (in particular, a currency unit), so a new Units package was derived and is included in OpenMDAO as openmdao.units. This package has the same basic function as that of PhysicalQuantities, but to make it more extensible, the unit definitions were moved from the internal dictionary into an externally readable text file called unitLibdefault.ini. See the source documentation for more information on the OpenMDAO Units package, including how to add units.

As an example, consider a component that calculates a pressure (in Pascals) given a known force (in Newtons) applied to a known area (in square meters). Such a component would look like this:

from openmdao.main.api import Component
from openmdao.lib.datatypes.api import Float

class Pressure(Component):
    """Simple component to calculate pressure given force and area"""

    # set up interface to the framework
    force = Float(1.0, iotype='in', desc='force', units='N')
    area = Float(1.0, iotype='in', low=0.0, exclude_low=True, desc='m*m')

    pressure = Float(1.0, iotype='out', desc='Pa')

    def execute(self):
        """calculate pressure"""

        self.pressure = self.force/self.area

The low and exclude_low parameters are used in the declaration of area to prevent a value of zero from being assigned, resulting in a division error. Of course, you could still get very large values for pressure if area is near machine zero.

This units library can also be used to convert internal variables by importing the function convert_units from openmdao.main.api.

>>> from openmdao.main.api import convert_units
>>> convert_units(12.0,'inch','ft')

Coercion and Casting

OpenMDAO variables have a certain pre-defined behavior when a value from a variable of a different type is assigned. Variables were created using the casting traits as opposed to the coercion traits. This means that most mis-assignments in variable connections (e.g., a float connected to a string) should generate a TraitError exception. However, certain widening coercions are permitted (e.g., Int->Float, Bool->Int, Bool->Float). No coercion from Str or to Str is allowed. If you need to apply different coercion behavior, it should be simple to create a Python component to do the type translation.

More details can be found in the Traits 3 User Manual.

Variable Containers

For components with many variables, it is often useful to compartmentalize them into a hierarchy of containers to enhance readability and “findability.”

Variables in OpenMDAO can be compartmentalized by creating a container from the Container base class. This container merely contains variables or other containers.

Normally a variable is accessed in the data hierarchy as:


but when it is in a container, it can be accessed as:


Consider an example of an aircraft simulation that requires values for three variables that define two flight conditions:

from openmdao.main.api import Component, Container
from openmdao.lib.datatypes.api import Float

class FlightCondition(Container):
    """Container of variables"""

    airspeed = Float(120.0, iotype='in', units='nmi/h')
    angle_of_attack = Float(0.0, iotype='in', units='deg')
    sideslip_angle = Float(0.0, iotype='in', units='deg')

class AircraftSim(Component):
    """This component contains variables in a container"""

    weight = Float(5400.0, iotype='in', units='kg')
    # etc.

    def __init__(self):
        """Instantiate variable containers here"""

        super(AircraftSim, self).__init__()

        # Instantiate and add our variable containers.
        self.add('fcc1', FlightCondition())
        self.add('fcc2', FlightCondition())

    def execute(self):
        """Do something."""

        print "FCC1 angle of attack = ", self.fcc1.angle_of_attack
        print "FCC2 angle of attack = ", self.fcc2.angle_of_attack

Here, the container FlightCondition was defined, containing three variables. The component AircraftSim is also defined with a variable weight and two variable containers fcc1 and fcc2. We can access weight through self.weight; likewise, we can access the airspeed of the second flight condition through self.fcc2.airspeed. We can also add containers to containers.

An interesting thing about this example is that we’ve implemented a data structure with this container and used it to create multiple copies of a set of variables. This can prove useful for blocks of variables that are repeated in a component. At the framework level, connections are still made by connecting individual variables. It is possible to create a custom data structure that the framework sees as a single entity for connection purposes. This is explained in Building a Variable Plugin from a Python Module.

Building a Simulation Model

A model is a hierarchical collection of components with an assembly at its root. The root assembly is also called the top level assembly. Executing the top level assembly executes the entire model.

Consider the top level assembly that was created for the simple tutorial problem.

from openmdao.main.api import Assembly
from openmdao.lib.drivers.api import CONMINdriver
from openmdao.examples.simple.paraboloid import Paraboloid

class OptimizationUnconstrained(Assembly):
    """Unconstrained optimization of the Paraboloid with CONMIN."""

    def __init__(self):
        """ Creates a new Assembly containing a Paraboloid and an optimizer"""

        super(OptimizationUnconstrained, self).__init__()

        # Create CONMIN Optimizer instance
        self.add('driver', CONMINdriver())

        # Create Paraboloid component instances
        self.add('paraboloid', Paraboloid())

        # Add to driver's workflow

We can see here that components that comprise the top level of this model are declared in the __init__ function. The base class __init__ function is called (with the super function) before anything is added to the empty assembly. This is important to ensure that internal framework machinery has been properly initialized before any methods such as add are called.

The add method takes a valid OpenMDAO name and a corresponding component instance as its arguments. This function call adds the instance to the OpenMDAO model hierarchy using the given name. In this case then, the CONMIN driver is accessible anywhere in this assembly via self.driver. Likewise, the Paraboloid is accessed via self.paraboloid.

A Component can also be removed from an Assembly using remove.


An Assembly is a special type of Component with the characteristics below. It contains:

  • Some number of other components (some of which may be assemblies)
  • At least one Driver with the name driver. Each Driver has its own workflow.

An Assembly retains the Component API (i.e., it can be executed, added to models, and exists in the model hierarchy), but it also extends the API to include functions that support the above-listed characteristics.

Connecting Components

Consider once again the top level assembly that was created for the simple tutorial. We would like to create a few instances of the Paraboloid function and connect them together in series.

from openmdao.main.api import Assembly
from openmdao.examples.simple.paraboloid import Paraboloid

class ConnectingComponents(Assembly):
    """ Top level assembly for optimizing a vehicle. """

    def __init__(self):
        """ Creates a new Assembly containing a Paraboloid and an optimizer"""



Components are connected by using the connect function built into the assembly. Connect takes two arguments, the first of which must be a component output, and the second of which must be a component input. These are expressed using their locations in the OpenMDAO model hierarchy with respect to the scope of their parent assembly. Additionally, only one output can be connected to any input. On the other hand, it is fine to connect an output to multiple inputs. The violation of any of these rules raises an exception.

A variable is not required to be connected to anything. Typical components will have numerous inputs, and many of these will contain values that are set by the user or are perfectly fine at their defaults.

Variables can be added to an assembly and used to promote internal variables, making them visible to components outside of the assembly. There is a convenience function called create_passthrough that creates a variable in the assembly and connects it to an internal component variable in one step.

Consider a similar assembly as shown above, except that we want to promote the remaining unconnected variables to the assembly boundary so that they can be linked at that level.

from openmdao.main.api import Assembly
from openmdao.examples.simple.paraboloid import Paraboloid

class ConnectingComponents(Assembly):
    """ Top level assembly for optimizing a vehicle. """

    def __init__(self):
        """ Creates a new Assembly containing a Paraboloid and an optimizer"""

        super(ConnectingComponents, self).__init__()




The create_passthrough function creates a variable on the assembly. This new variable has the same name, iotype, default value, units, description, and range characteristics as the original variable on the subcomponent. If you would like to present a different interface external to the assembly (perhaps you would like different units), then a passthrough cannot be used. Instead, the desired variables must be manually created and connected. You can find a more detailed example of this in the complex tutorial. Most of the time passthroughs are sufficient.

Assemblies also include a way to break variable connections. The disconnect function can be called to break the connection between an input and an output or to break all connections to an input or output.

>>> from openmdao.examples.enginedesign.vehicle import Vehicle
>>> my_car = Vehicle()
>>> # Disconnect all connections to tire_circumference (total:2)
>>> my_car.disconnect('tire_circumference')
>>> # Disconnect a specific connection
>>> my_car.disconnect('velocity','transmission.velocity')

You probably won’t need to use disconnect very often. However, some components may need to reconfigure their connections during runtime, so it is available.

Interacting with Files and Directories

Many components will need to read from and write to files during model execution. For example, a component might need to generate input files for and parse output files from an external application. In order to write components such as these, it is important to understand how objects in OpenMDAO interact with the file system.

The top assembly in the OpenMDAO model hierarchy contains the root path. This path is not known until after the assembly is instantiated (to learn how to set the root path, see Setting the Top Level Assembly). All components that are part of an assembly with a valid absolute directory have the same absolute directory.

You can change the absolute path of the working directory for any component on instantiation by setting the directory attribute in the __init__ function. For example, given the simple optimization model, we can specify a new working directory for the Paraboloid component when it is instantiated.

from openmdao.main.api import Assembly
from openmdao.lib.drivers.api import CONMINdriver
from openmdao.examples.simple.paraboloid import Paraboloid

class OptimizationUnconstrained(Assembly):
    """Unconstrained optimization of the Paraboloid with CONMIN."""

    def __init__(self):
        """ Creates a new Assembly containing a Paraboloid and an optimizer"""

        super(OptimizationUnconstrained, self).__init__()

        # Create Paraboloid component instances
        self.add('paraboloid', Paraboloid(directory='folder/subfolder'))

Notice that this is a relative path. All components in the model hierarchy must operate in a directory that is a sub-directory of the top level assembly’s absolute path. If you attempt to give a component an absolute path that is not a descendant of the top assembly’s absolute path, OpenMDAO will terminate with an exception. If two components need to operate in directories disparate from the top path in the hierarchy (e.g., one component in the model needs to run on a scratch disc), then this can be accomplished by using multiprocessing, wherein each process has its own top level.


Drivers are generally iterative solvers, such as optimizers, that operate on their respective workflow until certain conditions are met. OpenMDAO includes several drivers that are distributable (i.e., either open source or public domain.) This section describes the driver interface that is common to most drivers. A more complete discussion on how to use each of the drivers can be found in the section on Drivers in Appendix B: Standard Library Reference.

The Driver API

Drivers in OpenMDAO share a functional interface for setting up certain common parts of the problem. There are functions to handle parameters, which are inputs to a system and are also known as design variables for optimizers or independents for solvers. Likewise, there are also functions to handle constraints.

To illustrate the parameter interface, consider a model in which our goal is to optimize the design of a vehicle with several design variables using the CONMINdriver optimizer.

from openmdao.main.api import Assembly
from openmdao.lib.drivers.api import CONMINdriver

class EngineOpt(Assembly):
    """ Top level assembly for optimizing a vehicle. """

    def __init__(self):
        """ Creates a new Assembly containing a DrivingSim and an optimizer"""

        super(EngineOptimization, self).__init__()

        # Create DrivingSim component instances
        self.add('driving_sim', DrivingSim())

        # Create CONMIN Optimizer instance
        self.add('driver', CONMINdriver())

        # add DrivingSim to workflow

We add design variables to the driver self.driver using the add_parameter function.

# CONMIN Design Variables
self.driver.add_parameter('driving_sim.spark_angle', low=-50. , high=10.)
self.driver.add_parameter('driving_sim.bore', low=65. , high=100.)

Parameters are assigned via a string that contains the pathname of an OpenMDAO variable. This variable must exist in the scope of the assembly that contains the driver. In other words, if an assembly contains a driver, the parameters added to that driver cannot be located outside of that assembly. Also, each parameter must point to a component input, not a component output. During driver execution, the parameter values are set, and the relevant portion of the model is executed to evaluate the new objective.

The low and high arguments can be used to specify an allowable range for a parameter. Using these parameters is useful for optimization problems where the design variables are constrained. Generally, the optimizer treats these as a special kind of constraint, so they should be defined using the low and high parameters rather than the add_constraint method. If low and high values are not given, then they are pulled from the corresponding low and high parameters that are defined in the variable. If low and high aren’t defined in either place, then an exception is raised. Some drivers (in particular solvers) do not support a low or high value; in such a case, you can just set each of them to a large number, e.g., low=-1e99 and high=1e99.

Multiple parameters can also be added in a single call to add_parameters (note the letter s) by passing a list of tuples.

# Some more Design Variables
self.driver.add_parameters([ ('driving_sim.conrod', 65.0 , 90.0),
                             ('driving_sim.IVC', 0.0, 90.0) ])

The IHasParameters interface also includes some other functions that are more useful when used interactively or when writing more advanced components. The functions list_parameters, remove_parameters, and clear_parameters can be used to respectively list all parameters, delete a single parameter, and clear all parameters.

>>> from openmdao.examples.simple.optimization_constrained import OptimizationConstrained
>>> top = OptimizationConstrained()
>>> top.driver.list_parameters()
['paraboloid.x', 'paraboloid.y']
>>> top.driver.remove_parameter('paraboloid.x')
>>> top.driver.list_parameters()
>>> top.driver.clear_parameters()
>>> top.driver.list_parameters()

There are also get_parameters and set_parameters methods, but these methods are typically used by drivers to manage the parameters in their workflow and are not called directly by users. These will be described in the section Adding new Drivers.

A similar interface is present for interacting with constraints. Constraints are defined using strings containing equations or inequalities that reference available OpenMDAO variables. Both equality and inequality constraints are supported via the interface; however, when you use a driver, you should verify that it supports the desired type of constraint. For example, the CONMIN driver supports inequality constraints but not equality constraints.

Constraints are added to a driver using the add_constraint method.

self.driver.add_constraint('driving_sim.stroke < driving_sim.bore')

Constraints are defined using boolean expressions, so they are considered to be satisfied when the expressions evaluate to True and violated when they evaluate to False. The following constraint declarations are all equivalent:

self.driver.add_constraint('driving_sim.stroke - driving_sim.bore < 0')
self.driver.add_constraint('driving_sim.stroke < driving_sim.bore')
self.driver.add_constraint('driving_sim.bore > driving_sim.stroke')

Using the eval_eq_constraints and eval_ineq_constraints methods, an optimizer or solver can query for the status and values of its constraints. Both methods return a list of tuples of the form (lhs, rhs, relation, result), where lhs is the value of the left hand side of the expression, rhs is the value of the right hand side of the expression, result is the boolean result of evaluating the expression, and relation is a string indicating the type of relation used in the expression, e.g., >, <, >=, <=, or =. The values of the left- and right-hand sides are needed by gradient optimizers that apply the constraint via a penalty function.

The IHasConstraints interface also supports equality constraints. At present, none of the optimizers in OpenMDAO support equality constraints, but they are used by the BroydenSolver to assign the dependent equation. The syntax includes an equal sign in the expression.

self.driver.add_constraint('dis1.y1 = 0.0')


OpenMDAO does not check for duplicate constraints, so be careful when adding them.

Sometimes you want to change the scaling on constraints, particularly for cases where the constrained variables are of disparate orders of magnitude. You can do this conveniently with the optional scale argument in the call to add_constraint.

self.driver.add_constraint('driving_sim.stroke - driving_sim.bore < .00001', scaler=10000.0)

Here, the constraint has been scaled up so that when its value is passed to the optimizer, it is in a similar range (and hence, of similar weight) as the other constraints in the model. Although an optional adder argument was also added to shift both the left- and the right-hand sides of a constraint, the current OpenMDAO gradient optimizer (CONMINdriver) internally shifts all constraints to the origin, so this parameter is not needed.

Constraints can be removed using remove_constraint. The same string used to add the constraint should be used to remove it. Whitespace within the expression is ignored.

self.driver.remove_constraint('dis1.y1 = 0.0')

A list of constraint expression strings can be obtained using list_constraints.

lst = self.driver.list_constraints()

Calling clear_constraints will remove all constraints from a driver.


Finally, OpenMDAO uses a similar interface for specifying objectives. A single objective (some future optimizers will handle multiple objectives) can be added to a driver using the add_objective method with an argument that is a string expression built up from available OpenMDAO outputs.

# CONMIN Objective = Maximize weighted sum of EPA city and highway fuel economy
self.driver.add_objective('-(.93*driving_sim.EPA_city + 1.07*driving_sim.EPA_highway)')

In this example, the objective is to maximize the weighted sum of two variables. The equation must be constructed using valid Python operators. All variables in the function are expressed in the scope of the local assembly that contains the driver.

For drivers that only operate on a single objective (e.g., CONMIN), you can replace the current objective by calling add_objective with the new objective as an argument.

# Replace the objective with EPA_highway

The IHasObjective interface also includes functions to list the objective and to query for the objective value.

>>> from openmdao.examples.simple.optimization_unconstrained import OptimizationUnconstrained
>>> model = OptimizationUnconstrained()
>>> model.driver.list_objective()
>>> model.driver.eval_objective()

Adding new Drivers


Show how to add new drivers.

Running OpenMDAO

Setting the Top Level Assembly

When a Component or Assembly is instantiated as a standalone object, it is not aware of the directory where it resides. Any component added to such an assembly also does not know its path. The function set_as_top is available to denote an assembly as the top level assembly in the framework. Once an assembly is set as the top level assembly, it gains an absolute path which can be accessed through the function get_abs_directory.

The path that is set by set_as_top is always the current working directory in the Python environment.

>>> from openmdao.main.api import Assembly, set_as_top
>>> z1 = Assembly()
>>> z1.get_abs_directory()
Traceback (most recent call last):
RuntimeError: can't call get_abs_directory before hierarchy is defined
>>> set_as_top(z1)
<openmdao.main.assembly.Assembly object at ...>
>>> z1.get_abs_directory()

The output in this example depends on your local directory structure. All components added into this assembly will have this same absolute path. If a component or assembly does not have a valid absolute directory, then File variables will not be able to read, write, or even open their target files.

Executing Models


Show how to run a model.


Discuss Reset to Defaults.

Error Logging & Debugging


Explain the error logging capability.

Saving & Loading


Show how to save and load.

Sharing Models


Discuss sharing models.


The execution order for components in a model is determined by the workflow object that the components belong to. OpenMDAO current has two available workflow classes that are described below. They are Dataflow and SequentialWorkflow.


The “default” workflow for a model is inferred from the data flow connections. This means that a component is available to run once its inputs become valid, which occurs when the components that supply those inputs are valid. Since direct circular connections (algebraic loops for those familiar with Simulink) are not permitted, there will always be an execution order that can be determined from the connections. In the absence of a connection between two components, this workflow will attempt to execute them in the order that they were added to it.

When any component input is set, all dependent outputs are invalidated. If an input is connected to an output and that output becomes invalid, then the input is also invalid. If a component has any invalid inputs or outputs, it will be executed during the next run. When a component’s inputs are changed, all downstream variables that depend on them either directly or indirectly are invalidated. Also, when a model is instantiated, all outputs are invalid, which ensures that the whole model always executes the first time it is run.


This workflow is a simple sequence of components. The components will be executed in the order that they were added to the workflow regardless of data dependencies. Generally, this is a bad idea, but it’s here for those rare occasions when the exact sequence must be specified.

Geometry in OpenMDAO

We are currently investigating an API to provide a unified geometry interface. More information on the notional prototype can be found in Geometry Interfaces in OpenMDAO (Working Document).