Exercise: Monte Carlo

Monte Carlo methods may be familiar from earlier years, where stochastic (random) processes are used to find, for example, the minimum energy position of some atom in space. Here we will look at using Monte Carlo to improve our model-fitting methods, in particular the quantification of the uncertainties on the model parameters.

In the model fitting section, you were shown how the curve_fit function can get estimations of these inverse uncertainties (a fancy name for the parameter uncertainties in the model). However, these uncertainties are only estimates and are based on assumptions about the fitting process (the particular assumptions made are beyond the scope of this course). Therefore, we will use a stochastic method to more completely probe these uncertainties.

For this problem, we are going to look at the investigation of a mixture of organic species by IR spectroscopy. We have been given this data set, and are told it contains a mixture of toluene and benzyl alcohol.

Use the !head command to investigate the experimental data file and plot the data using the plt.errorbar function.

Our model

This exercise aims to determine the relative concentration of toluene and benzyl alcohol that make up the mixture data that you have just plotted. To achieve this, we must define a model to describe this data. The transmittance data from the mixture (\(T_{\text{mix}}\)) is made up of transmittance data from toluene (\(T_{\text{t}}\)) and benzyl alcohol (\(T_{\text{b}}\)), and nothing else. Since we are interested in the relative concentration, we can describe this as a fraction of the composition and since the transmittances are additive, we can define the following model.

\[T_{\text{mix}} = cT_{\text{t}} + (1 - c)T_{\text{b}}, \]

where \(c\) describes the composition of the mixture and is the single parameter in our model.

Write a function for the above model, c should be the first argument as this is the variable parameter we are investigating. Other arguments should include the transmittance from the pure samples.

We don’t yet have IR spectra for the pure samples that we can use in the above model. Files for these spectra are available here for toluene and here for benzyl alcohol.

Read in and plot each of the pure datasets.

Interpolation

You will have noticed that the x-axis on each of the three plots created so far do not cover the same range. However, to accurately use the model outline above, the transmittance values for all three IR spectra should be at the same wavenumbers. Therefore, we must interpolate the two pure datasets such that the wavenumber values are the same as for the mixture.

Interpolation is where we determine new data points within the range of a discrete set of known points. Essentially, we use what we know about the x- and y-values to determine the y-values for a different set of x-values. It is important that the new range of x-values is from within the existing range, or else we are extrapolating (which is often unscientific). In the three datasets that you have, the two pure samples both have broader ranges than the mixture. Therefore, we will use the x-values for the mixture and extrapolate new y-values for each of the pure samples.

For interpolation, we can use the np.interp method. This method takes three arguments, x the new x-axis, xp the old x-axis, and fp the old y-axis. It will return a new set of y-values.

Interpolate transmittance values for the two model IR spectra and plot these over the original data to check that they agree.

Optimisation

Now that we have the pure data read in and on the correct x-axis, we can test the model that we created earlier.

Generate the model transmittance data that would arise from a 50:50 mixture of the two components. Plot this data on top of the mixture data and see if it agrees visually.

Now that our model is working, we can use the methodology introduced previously to minimise the difference between our model and the data.

Write a chi-squared function and minimise this using the scipy.optimize.minimize function. This will give an optimised value for c.

Sampling

Having found the optimised value for the concentration, we can now use a modification of a Monte Carlo process to sample the uncertainty in this concentration parameter. This methodology is called Markov chain Monte Carlo (MCMC), it involves starting from some value and then changing the value by some small random amount with each iteration. The fact that the next parameter value depends on the previous one makes the sampling a Markov chain, while the use of random perturbations is Monte Carlo.

In MCMC we start from an initial value, usually found by obtaining an optimised solution. This initial value is changed by some random amount (\(\delta\)), which is obtained based on some defined. step size (\(s\)) change with respect to the variable (\(v\)),

\[ \delta = Nsv, \]

where \(N\) is some random number, obtained from a normal distribution centred on 0 with a standard deviation of 1. In Python, \(R\) can be obtained with the function np.random.randn(). We then determine if this random perturbation has improved the. agreement to the data or not. If it has, we accept this new value for our variable (\(v + \delta\)) and perform another perturbation. If this perturbation does not improve agreement with the data, the new value is not immediately rejected, rather it is only rejected if the probability of this transition (\(p\)) is less than some random number from 0 to 1 (this time we use np.random.random() to obtain such a number). The probability is found by,

\[p = \exp\Bigg(\frac{-\chi^2_{\text{new}} + \chi^2}{2}\Bigg), \]

where \(\chi^2\) is the original goodness of fit and \(\chi^2_{\text{new}}\) is the goodness of fit after the perturbation. This means that it is possible for the agreement to get worse over time. However, the amount by which it can get worse is controlled by the probability. The result of this is that the values of our Markov chain that are accepted will describe the statistically feasible value for our parameter given the uncertainty in the experimental measurements.

The algorithm for a typical MCMC sampling process is as follows:

  1. Create an empty list for the accepted values

  2. Evaluate \(\chi^2\) for the initial guess, typically this initial guess will be the optimised solution.

  3. Perturb the parameter value (\(v + \delta\))

  4. Calculate \(\chi^2_{\text{new}}\) for the perturbed value

  5. Determine the probability of this transition

  6. Check that \(p\geq R\), where \(R\) is a random number from 0 to 1, if the new \(\chi^2\) is less than the old one, then \(p>1\) and therefore always accepted

  7. If true, updated the values of \(v\) and \(\chi^2\), and append \(v\) to the list of the accepted values

  8. Go to step 3 and repeat until the desired number of iterations have been achieved.

Write a function to perform the MCMC algorithm outlined above, this should take a number of iterations (this should be no more than 2000) and step size as arguments and return the list of accepted values. Plot a histogram of the accepted values (using plt.hist), these should be normally distributed (you may need to google to see what this looks like). Vary the step size between 1 and 0.001 to get the most normally distributed data you can. Using the statistical functions in NumPy, calculate the mean and standard deviation of this distribution.

Worked Example