Figure 5.

Ablation analyses by removing planar information. Direct comparison of CFFN with FFN, a variant without 2D information, shows that the former is more accurate, robust, data-efficient and error-proof in different training samples (500,250,150,100 samples respectively) and flawed data proportions (10% or 20%). The baseline is 500 training samples. Three types of flaws are: type 1, missing one atomic coordinate; type 2, missing one atomic force; and type 3, inaccurate molecular energy. Units of MAE are in kcal^*mol⁻¹ Å⁻¹.