Physics-Informed Loss#
Function#
The learned function, \(\boldsymbol{f_\theta}\), will map the spatial coordinates, \(\mathbf{x}_\phi \in \mathbb{R}^{D\phi}\), and time coordinate, \(t \in \mathbb{R}\), to sea surface height, \(u \in \mathbb{R}\).
Loss#
The standard loss term is data-driven
However, there is no penalization to make the field behave the way we would expect. We also want a regularization which makes the field, \(u\), behave how we would expect. This can be achieved by adding a physics-informed loss regularization term to the total loss.
This loss term can be minimized by effectively minimizing a PDE function. For example:
where \(\partial_t\) is the derivative of the field, \(u\), wrt to time and \(\mathcal{N}[\cdot]\) are some partial differential equations. We are interested in minimizing the full PDE, which we denote \(\boldsymbol{f}_{phy}\), st it is 0. So the standard loss function applies.
Examples#
Below are some examples of how I have used the PINNs loss/regularization formulation in my own research.
QG Equation#
We have the following PDE for the QG dynamics:
where \(q(x,t) \in \mathbb{R}^2 \times \mathbb{R} \rightarrow \mathbb{R}\) is the potential vorticity (PV), \(\psi(x,t) \in \mathbb{R}^2 \times \mathbb{R} \rightarrow \mathbb{R}\) is the stream function, \(\partial_t\) is the partial derivative wrt \(t\), \(\boldsymbol{J}\), is the Jacobian operator and \(\det \boldsymbol{J}(\cdot,\cdot)\) is the determinant of the Jacobian.
Objective: We want to convert this PDE in terms of sea surface height (SSH) instead of PV and the stream function.
QG Equation 4 SSH (TLDR)#
Note: For the ease of notation, let’s denote \(u\) as the SSH. The above PDE can be written in terms of \(u\)
See the following page for more and how this equation was derived.