Performance of three trained NeuralIL models on distinct test data sets (5000 4×1 structures, 1100 5×1 structures, 149 GPAW-optimized 5×1 structures), comparing the MAEs of the force components and of the energies per surface area.

Test set	MAE	NNFF1	NNFF2	NNFF3
4×1	f/meV Å−1	77.19	84.22	79.76
	E/meV Å−2	1.35	2.07	1.61
5×1	f/meV Å−1	278.55	123.46	126.37
	E/meV Å−2	9.89	1.92	2.21
Opt	f/meV Å−1	175.40	105.35	107.66
	E/meV Å−2	8.95	7.72	6.59