what kind of algorithm are you using? it looks like your have set some boundary conditions/restrictions for the fitting parameters (ranges for center wavenumber, line width, etc), which do not allow a better fitting.

the signal-to-noise ratio is quite good. it may be an option to do a true mathematical deconvolution via the derivative spectra

I apologize, I completely forgot to add that information. I am using Gaussian equation to determine secondary peaks. The only parameters I have set are the location of the center of the peaks. I have set 1656 (a helix) 1642 (b sheet) , 1624 (b sheet) and 1648 (random coil). 

have you tried the fit without setting the center wavenumber? for example, just looking at the data it seems that there is a a shoulder band around 1690 cm-1.

one of the profiles (centered 1624?) seems to be extremely broad and represent more a background/baseline than a contribution to the band. have you tried a baseline correction before the fitting?

Mr. Bullard,

In your question there is mixed three terms in the IR-quantitative spectroscopy, furthermore wrongly, because of:

1. Deconvolution is a procedure applicable to determination of the number of the sub-maxima in your overlapped IR-pattern. Please pay attention to the attachment Fig. 1. It consists of keeping of the integral intensity under the curve and reducing the peak-half-width.

2. Once when you have obtain the number of sub maxima (in the shown as attachment case, there are 5, you can apply curve-fitting as shown by Figs. 2 and 3. If you pay attention to both the figures, when you have used number of peaks 5, your chemometric analysis have yielded to a better confidence level (99.91 %).

3. You have shown in your figures as intensity the band-maximum on the ordinate, but this is not the intensity value of a IR – measurements. Because of such as experimental analysis is conducted via recording A=f(ni) or T=f(ni). In this context when you operate with dataset measured as A=f(ni), you should bear in mind that the absolute intensity is determined by Eq.(1) (attachment). In fact this is the Einstein coefficient, where to the transition dipole moment is related the integral absorption intensity (it is given by Eq.(3)). In this equation e(ni) is molar extinction coef., which however is joined with the absorptivity (or the absorption coefficient in a IR – spectrum) via Eq.(2). Given that if you would like to know the absolute intensity, this means that you should operate with the integrated value of the (ni)th band. Please pay attention to Figs. 3 and 4. If you have decided to obtain the intensity of the A5 band (3730 cm-1), you even cannot determine the peak position (or lmax) without curve-fitting. The relation A5/Atot = 0.0732 (Atot – value obtained by integration of the whole multi-band spectroscopic profile). This value is different by A*, however, which you would like to take as intensity. Taking this A* value, you do not account for peak-half-width. But like UV–VIS spectroscopic profile, the IR-bands are also relatively broad and, thus they cannot be characterized only by ni0, but by peak-half-width as well.

Details about point 3 you may find in series refs, for example the shown ones:

[Refs.] (a) A. Jolly, Y. Benila, A. Fayt, New infrared integrated band intensities for HC3N and extensive line list for the m5 and m6 bending modes, Journal of Molecular Spectroscopy 242 (2007) 46–54;(b) Arthur S. Wexler, Integrated Intensities of Absorption Bands in Infrared Spectroscopy, Applied Spectroscopy Reviews, 1:1 (1967) 29-98; (c) G.Lebron, T.Tan, Integrated Band Intensities of Ethylene (12C2H4) by Fourier TransformInfrared Spectroscopy, International Journal of Spectroscopy, Volume 2012, Article ID 474639, 5 pages;(d)H. William Wilson, Integrated Infrared Band Intensities, Journal of Chemical Educalion, 392;(e) H. Hagemam, R.Snyder, A.Peacock, L. Mandelkern, Quantitative Infrared Methods for the Measurement of Crystallinity and Its Temperature Dependence: Polyethylene, Macromolecules 1989, 22, 3600-3606

In any curve fitting, if you have more 'curves' in your Gaussian modeling than there are actual peaks, then you have more degrees of freedom than you have peaks, and you'll overfit the data, leading to a very unstable calculation.  The usual way around this is to start with the minimum number of peaks justifiable based on what you see on the spectrum.  Another way (since there may be lots of peaks in actuality and yet multiple combinations of curves to fit the data!) is to constrain each of your curves -- GRAMS software for example allows you to constrain peak position and even band shape (I think you could vary the mix of Lorentzian and Gaussian, if you think that is justifiable to use).  So that would allow you to reach a curve fit solution that is stable given those constraints.  One more constraint that is relevant in your case is to force all intensities to be positive -- in your 3rd graph, there is a 'negative' curve as a component in your fit, which is nonsensical in most cases (except maybe when you're doing a spectral hole burning experiment).

In conclusion, any time you have unstable curve fitting, suspect 1. too many peaks, and/or 2. not enough constraints applied to your Gaussian curves.