# Multifidelity modeling#

We will discuss multifidelity modeling, which is a technique for constructing surrogate models that combine information from multiple sources of varying fidelity. The goal is to leverage the strengths of each fidelity level to build a more accurate and efficient surrogate model.

## Examples of Multifidelity Information#

You could have a finite element model of a structure that is very accurate but computationally expensive. This would be your high fidelity model. You could also have a simplified analytical model that is much faster to evaluate but less accurate. This would be your low fidelity model.

You could have a finite element model with a fine mesh (high fidelity) and a coarse mesh (low fidelity).

You could have experimental data (high fidelity) and a simplified analytical model (low fidelity).

and so on

## Notation#

Let \(f_l\) denote the low-fidelity and \(f_h\) denote the high-fidelity model. The expectation is that you have a lot of data from the low-fidelity model and only a few data points from the high-fidelity model. Suppose that our dataset is as follows:

\(\mathbf{X}_l = \{x_{l,i}\}_{i=1}^{n_l}\) and \(\mathbf{y}_l = \{y_{l,i}\}_{i=1}^{n_l}\) are the training data from the low-fidelity model.

\(\mathbf{X}_h = \{x_{h,i}\}_{i=1}^{n_h}\) and \(\mathbf{y}_h = \{y_{h,i}\}_{i=1}^{n_h}\) are the training data from the high-fidelity model.

To keep the notation simple, we will not be showing \(\mathbf{X}_l\) and \(\mathbf{X}_h\) in the equations below. But whenever \(\mathbf{y}_l\) or \(\mathbf{y}_h\) appears, it is assumed that the corresponding \(\mathbf{X}_l\) or \(\mathbf{X}_h\) is also present.

## What we are after#

We are after the best thing we can say about the high-fidelity model given all the data. That is, formally we are after this:

To be abl to pull this off, we need two ingredients:

Regression for the low-fidelity model: \(p(f_l|\mathbf{X}_l, \mathbf{y}_l)\).

A model that relates the low-fidelity and high-fidelity models: \(p(f_h|f_l, \mathbf{X}_l, \mathbf{y}_l, \mathbf{X}_h, \mathbf{y}_h)\).

## Regression for the low-fidelity model#

We are just going to do Gaussian process regression for the low-fidelity model. This is the easy part. We start with a prior:

Assume that the observations are noisy:

We then condition on the data to get the posterior over \(f_l\), which is also a Gaussian process:

Here the posterior mean function is:

and the posterior covariance function is:

where as usual \(I\) is the identity matrix, and \(k_l(x, \mathbf{X}_l)\) is the covariance between \(x\) and the training data, and \(k_l(\mathbf{X}_l, \mathbf{X}_l)\) is the covariance between the training data.

## Relating the low-fidelity and high-fidelity models#

The simplest way to relate the low-fidelity and the high-fidelity models is to assume that the high-fidelity model is just the low-fidelity model plus some discrepancy:

where \(\rho\) is a scaling factor and \(\epsilon\) is a discrepancy term. The scaling factor is typically assumed to be a scalar, but it could also be a function of the input, e.g.,

where \(w_i\) are weights and \(\phi_i(x)\) are basis functions.

Kennedy and O’Hagan, 2000 suggest that the discrepancy term is a Gaussian process:

They also assume that this discrepancy term is independent of the low-fidelity model, formally:

Let’s work out the posterior over the high-fidelity model given the low-fidelity model. It is just:

where the mean function is:

and the covariance function is:

Now, let’s condition on the low-fidelity data and integrate out the low-fidelity model \(f_l\). Let’s start with the mean:

Now, let’s look at the covariance:

If \(\rho\) was a function of the input, then we would have:

and

So, we conclude that the high-fidelity model given the low-fidelity data is also a Gaussian process:

We can now condition this on the high-fidelity data to get the posterior over the high-fidelity model given all the data. We still get a Gaussian process:

The posterior mean is:

and the posterior covariance is:

We can fit the parameters by maximizing the marginal likelihood:

and each one of these terms is analytically tractable.

In practice, if we have a lot more low-fidelity data, we may first maximize \(\log p(\mathbf{y}_l)\) with respect to all low-fidelity model parameters, and then maximize \(\log p(y_h|\mathbf{y}_l)\) with respect to the high-fidelity model parameters.

## Another approach to low-fidelity and high-fidelity models#

Another approach is what we introduced in Karumuri et al., 2023. In this approach, assume that the high-fidelity model given the low-fidelity model is just a Gaussian process:

where the mean function is a standard user choice, e.g., a constant, a linear function, or a polynomial, but the covariance function is now:

where \(k\) is a kernel function that operates on the input and the low-fidelity model.

This is essentially a deep Gaussian process with one hidden layer. The high-fidelity covariance function does not just look at the input, but it also compares the similarity of the output low-fidelity models at the two input points.

Some work on general inference with this type of models has been done in Damianou et al., 2013. But it is not trivial to train such models. However, if we have a lot of low-fidelity data, we can simplfiy things quite a bit. Under the assumption that the posterior of the low-fidelity model collapses to the posterior mean, we can write:

and then take our kernel to be: