The whole idea behind regression is that there is error in measurements, and regression gives you the best fit through the data, minimizing the total error between all of the points and the fitted relationship. To get the best fit, you want a reasonable number of points spread evenly through the range of interest.
Your intention is a fit that goes exactly through each point, and this assumes that there's no variability in measurement and in other factors, which is highly unlikely. If I get some milk product with 3.25% butterfat and I make ten measurements, are all 10 going to give a reading of 14.528 on your sensor? What if the temperature varies slightly, or I measure before stirring, then after stirring, then after shaking the sample like crazy? What is the effect of different distributions of fat globule size? And how closely are you tring to measure butterfat? 3.25 ± 0.10? 3.25 ± 0.05? 3.25 ± 0.001?
Have you decided which is your independent variable (X) and which is your dependent variable (Y)? Regression assumes the values of the independent variable do not contribute to the error in the fit, so that the error is a combination of measurement error and the effect of other factors (noise factors, like temperature and so forth).
Also, whichever way you look at your points, you have two tightly spaced points with three other widely spaced points, and those two points have an inordinate influence on the fitted relationship, evidenced by the second relationship which fit very poorly.
What I would suggest is that you make a number of standard mixtures, more than five, evenly spread across the range of interest. Maybe 0 to 4% in increments of 0.5%. Have your samples at the same temperature as what your regular samples will be at, give them the same amount of agitation, make everything as near production measurement conditions as possible. Take your measurements. Repeat on another day. Look carefully at the results. Think hard, because it is hard to justify physically the use of a relationship beyond a first order fit. Don't use a statistical method to justify uning a nice curve that connects all of the points.
If what you want is just a smooth calibration curve, then draw it by hand. If you want it in Excel, then manually insert many points along your range where you think it looks nice, put the XY pairs into a lookup table, and use an interpolation formula (Google: Excel Interpolation) to get Y as a function of X.