Solving a physical problem requires simulation
Given parameters, \(\theta\), what is a configuration, \(x\), that satisfies the above equation?
Examples of \(\theta\):
Configuration could refer to the heat value at a particular time, or the position at a time value.
"I want to find the parameters, \(\theta\), that will admit the configuration, \(x\), that I observed/desire"
Examples include:
Each evaluation of the loss function requires a full simulation. To calculate a gradient, finite difference would require multiple runs of the same simulation.
PDEs are described in terms of function spaces. Consequently, the inverse problem is also in terms of function spaces rather than discrete vector spaces
The optimization will be effected by our choice of discretization
There's no one-size-fits-all solution!
We need \(d_{\theta}f\) in order to do any gradient based optimization
Derivative of dynamics
(application specific)
Derivative of cost functional
(closed form usually)
}
\(\lambda\) is known as the adjoint state
\(\lambda\) is the same dimension as the configuration, \(x\).
Solving for the adjoint state is no harder than running the forward simulation; it is often a linear solve
The adjoint state can be derived from Lagrange multipliers.
where \(x^*\) and \(\lambda^*\) are critical points.
One can view the adjoint state method as an extension of discrete Lagrange multipliers to functionals, where \(\lambda\) is a function instead of a coefficient.
In the PDECO literature, the KKT conditions are referenced a lot in the context of functional analysis.
From images of a volume, and I want to invert the image formation model to obtain the volume. Our equation of interest is the eikonal equation
Ground Truth
Initialization
Adjoint
Ground Truth
Initialization
Step 3
Final
Out of the box optimization gives good results!