Hi,
We have received several user reports in SciGlass Next about non-physical trends in temperature-dependent predictions, especially viscosity-related properties.
For example, for a fixed-composition homogeneous glass melt, viscosity should decrease monotonically with increasing temperature. Therefore, the characteristic viscosity temperatures should satisfy:
T1(logη=1) > T2(logη=2) > ... > T3(logη=13)
T1(logη=1) means Temperature for log viscosity η=1, η in Poise (P). It corresponds to T0 ( in Kelvin) in glasspy.
However, in some compositions with limited nearby training data, GlassNet can produce non-monotonic predictions.
I attached screenshots showing one example:
Composition:
{"name":["SiO2","ZnO","B2O3","Al2O3"],"value":[[25,44,25,6]]}For example, as shown below, the predicted value of T1 (logη=1) is quite low. For a homogeneous glass melt, it should be greater than T2 (logη=2), since viscosity decreases with increasing temperature.
After searching the database for this glass system, I found only a small amount of nearby data, which may be one of the reasons for the deviation:
(Only 14 entries for T1(logη=1) of this glass system.)
A similar trend problem also occurs with other temperature-dependent properties such as density at different temperatures.
Would it be possible to consider adding basic physical constraints, monotonicity checks, post-processing corrections, or at least warnings for temperature-dependent properties such as viscosity? What do you think?
Hi,
We have received several user reports in SciGlass Next about non-physical trends in temperature-dependent predictions, especially viscosity-related properties.
For example, for a fixed-composition homogeneous glass melt, viscosity should decrease monotonically with increasing temperature. Therefore, the characteristic viscosity temperatures should satisfy:
T1(logη=1) > T2(logη=2) > ... > T3(logη=13)
T1(logη=1) means Temperature for log viscosity η=1, η in Poise (P). It corresponds to T0 ( in Kelvin) in glasspy.
However, in some compositions with limited nearby training data, GlassNet can produce non-monotonic predictions.
I attached screenshots showing one example:
Composition:
{"name":["SiO2","ZnO","B2O3","Al2O3"],"value":[[25,44,25,6]]}For example, as shown below, the predicted value of T1 (logη=1) is quite low. For a homogeneous glass melt, it should be greater than T2 (logη=2), since viscosity decreases with increasing temperature.
After searching the database for this glass system, I found only a small amount of nearby data, which may be one of the reasons for the deviation:
(Only 14 entries for T1(logη=1) of this glass system.)
A similar trend problem also occurs with other temperature-dependent properties such as density at different temperatures.
Would it be possible to consider adding basic physical constraints, monotonicity checks, post-processing corrections, or at least warnings for temperature-dependent properties such as viscosity? What do you think?