In my simulation CaSO4:Dy TLD is inside a cassette having 3 discs as detector i.e. D1, D2 and D3. D1 is under (Al+Cu) filter of thickness 1.6 mm, D2 disc is under polystyrene of 1.6 mm and D3 is open (no filter). Parallel incident beams of energy 12.3-662 keV are incident one by one.
while placing it without ISO slab phantom, results of dose deposition in each disc is almost same in both codes i.e. Fluka and Geant4.
While placing it on ISO slab phantom (i.e. as in current input), back scattered part increase the dose as expected. But for energy < 70keV, energy deposited in D1 disc (i.e. under Al+Cu filter) in FLUKA is double of that in Geant4 while for other two discs both codes match the values.
Also dose in case of placing it on ISO slab Phantom is around 4-6 times without it in this energy range and i am not expecting that dose will increase to such a high value due to back scattering part.
I tried to use MULSPOT card to set single scattering but not much difference is observed.
At a first glance, the fact that the two detectors without filter (D2 and D3) give the same result with/without water phantom, while for the detector with filter (D1) results differ, may suggest that you’re witnessing the following (physical) boundary effect between (relatively) high Z and low Z materials:
Cu in the filter has a higher Z=29 than anything in the TLD (the highest you find there is Z=20 for Ca, which makes up about 40% in mass; the rest has much lower Z).
A higher rate of production of secondary electrons is expected in Cu than in the rest of nearby materials, the more so when you have the water phantom, since as you say, backscattering from it feeds the electron fluence (this may explain why without the water phantom the effect is not as noteworthy).
There will be more e- produced in the Cu filter that manage to make it into D1 than the other way around, leading to a larger dose deposition in D1.
This effect is of course e- transport/production-threshold dependent. To check the tentative explanation above, it may be sufficient to gradually raise the e- production/transport cuts and see if the magnitude of the effect decreases accordingly. Needless to say, the run with lower threshold will give the more accurate/realistic result.