You have a table with 100 values in terms of temperature. It's clearly nonlinear to some degree (not insanely so over that range).
So you have a few ways of proceeding. Generally what you want to do is to fit the data to a function f so you can find T = f(a) where a is a variable that goes from 0 to 1 for 0 to 100%.
There are many ways to go about that. One is to find a polynomial of some degree that minimizes the least-squares error for each point. If the power is less than 100 you'll have some error, and generally the whole thing will get more numerically unstable the higher the degree. You could try fitting a polynomial of reasonable degree (say up to about 8) and see how well it fits with the points and how it behaves between the points.
One method that uses more storage but is typically very numerically stable is to use cubic splines where the result is a smooth (differentiable) curve that passes exactly through each of the given points and behaves reasonably between them.
The numbskull simple method is to store the points and do linear interpolation rather than a cubic spline. That's like drawing a straight line between the points, so it's ugly but it hits all the points exactly and isn't too crazy between the points if there are enough and the curve isn't too nonlinear.
The optimum method will depend on the computing resources available and the required bandwidth and the required accuracy. Since the variable is temperature, the bandwidth is likely rather low, and since the numbers are only given to 4 digits probably the accuracy isn't all that demanding either.
MATLAB (curve fitting toolbox, ideally) which is very much not free, or one of the free programs with similar capabilities (Octave or Scilab) is an easy way to proceed and you can easily find how to do it with the search terms provided. You can even use Microsoft Excel in some cases.
Note that numerically poorly conditioned algorithms can have unexpected results since you end up coming with a result that subtracts almost-equal large numbers from each other, and truncation or rounding can affect the results cosmetically or even materially, so be sure to check the algorithm carefully before deploying it.