SAD versus PADS Input density

[hanaol]

I looked into two approaches for generating the input charge density:

1. Superposition of atomic densities (SAD)
2. Precomputed atomic density superposition (PADS)

For a small ~3K subset of the MP database the following table lists the CPU timing and the error metric for both the input and model (ResUNet) output.

Input type	CPU time (s/atom)	NMAE % (Input, MP)	NMAE % (UNet, MP)
SAD	1.4 (112 cores)	40.59	2.30
PADS	0.9 (1 core)	15.84	2.65

Here is a comparison of the learning and evaluation curves between SAD and PADS in W&B.

[hanaol]

After fixing the two points discussed in #104, I retrained the model on the subset and found an improved performance:

Method	NMAE % (Input, MP)	Train NMAE (%)	Validation NMAE (%)	Test NMAE (%)
SAD (NELM=0)	40.59	1.66	2.30	2.48
PADS (initial)	15.84	2.53	2.65	3.14
PADS (revised)	14.33	1.27	1.95	2.01

I then examined the error distribution across all splits and found no clear or systematic discrepancy between the input, prediction, and label. The main challenge instead appears to be a subset of highly localized or delocalized cases relative to the free-atom reference, which I will discuss in more detail later.

[forklady42]

Am I reading this correctly that the revised PADS is now performing better than the SAD input?