Table 1 DC3 benchmark dataset accuracy comparison between our classification approach and the methods tested in the DC3 work

	Al (fcc)	Fe (bcc)	Ti (hcp)	Si (cd)	H₂O (hd)	NaCl (sc)	Ar (fcc)	Li (bcc)	Mg (hcp)	Ge (cd)
a-CNA	50.3	39.9	15.7				57.1	34.4	47.3
PTM^a	95.9	84.3	82.8	99.9	100.0	94.6	96.9	83.1	95.7	99.2
Denoiser (1 step) + a-CNA	98.5	96.9	81.7				98.6	96.9	97.1
Denoiser (1 step) + PTM	100.0	99.5	98.7	100.0	100.0	99.3	100.0	99.6	99.9	100.0
Denoiser (8 steps) + a-CNA	100.0	99.7^b	100.0				100.0	100.0	100.0
Denoiser (8 steps) + PTM	100.0	99.9^b	100.0	100.0	100.0	100.0	100.0	100.0	100.0	100.0
DC3^a	96.9	86.8	89.4	99.0	99.2	95.6	97.5	85.8	97.4	100.0
i-CNA^a	68.5	56.6	27.9				75.7	51.6	64.9
AJA^a	66.9	35.6	42.4				74.0	34.1	67.0
VoroTop^a	24.0	61.2	57.5				23.1	57.2	57.4
Chill+^a				99.8	98.8					100.0

The accuracy value measured at the melting point, computed as the fraction of correctly labeled atoms, is shown in percentage. Our methods are shown in bold font. In our denoising approach, the RMSD (root-mean-square-deviation) cutoff for PTM is 0.1. Missing entries imply non-applicable structure types for the chosen classifier.
^aTaken from the DC3 paper²⁰.
^bThere is some trace amount of point defects in the BCC Fe snapshots at T = T_m (Supplementary Fig. 6b).

Quick links

Search