The method of using twilight flats to flat field the data suffers from several limitations.
Non-linear change in the intensity of the twilight sky with time
Since the sky is either decreasing in intensity (evening twilight flats) or increasing in intensity (morning twilight flats), since these changes are non-linear in time and since the array takes some time to read-out, one can get a discontinuity in the flat field between the two halves of the array which has an amplitude of 1% with the current DIT and NDIT settings.
The effect can be removed rather well by averaging the morning and evening twilight flats as the effect inverts itself between the two times.
Odd-even column effect
The odd even column effect should be removed from the individual flat fields. However, this only needs to be done for flat fields that were taken when the effect was strong. (See Sec. ).
The variable zero level offset
The zero level offset of the Hawaii array is a function of the flux, so when the flux of the array changes so does the offset. The change appears to be complex. As the flat is made from exposures of varying flux and since the zero level offset changes with flux, we are not measuring the relative sensitivity of a pixel, but more the relative sensitivity plus the change in the zero level offset.
At this point in time, we cannot remove the change in the zero level offset and this ultimately limits the accuracy at which the relative sensitivity of pixels and hence the flat field can be measured. Nevertheless, we can quantify the effect this has on the flat field by dividing one flat by another. The typical pixel to pixel RMS in the flat field is of the order of 6%. The typical pixel to pixel RMS of an image which is the division of one flat by another is 0.4-0.6%. This translates to an accuracy of 0.3-0.4% in the flat itself.
If the array is heavily saturated, two affects occur. Firstly, a remnant, whose timescale depends on the level of saturation, appears and, secondly, the QE of the array decreases. This has a much longer timescale, and appears to be a strong function of wavelength. One sees it strongest at J and weakest at Ks.
The solution to this problem is to not saturate the array, but this will occur from time to time, so flats need to be taken often enough so that the scientific data is not affected. Since the timescale of the effect is long, the current practice of taking flats as required is satisfactory.