Table 2.

Performance benchmarking for the prediction task of “materials property given structure”.

Dataset	Property	Model input	Size	Best of 10 ML	17-layer plain network	17-layer IRNet
JARVIS	Formation enthalpy (eV/atom)	Structure	25,405	0.125	0.153	0.114
		Structure+Comp	25,405	0.107	0.129	0.097
		Structure+Comp+EF	25,405	0.107	0.125	0.096
	Bandgap (eV)	Structure	22,952	0.338	0.400	0.337
		Structure+Comp	22,952	0.284	0.336	0.280
		Structure+Comp+EF	22,952	0.280	0.335	0.276
OQMD	Formation enthalpy (eV/atom)	Structure	435,582	0.106	0.102	0.072
		Structure+Comp	435,582	0.072	0.065	0.041
	Bandgap (eV)	Structure	435,293	0.056	0.075	0.071
		Structure+Comp	435,293	0.045	0.046	0.042
	Volume_pa (${\text{A}}^3$)	Structure	435,625	1.032	1.806	1.684
		Structure+Comp	435,625	0.451	0.385	0.287

The number in bold font represents the best model performance for a given combination of dataset, materials property and model input (each row).