Fig. 3: Average 3-body relevance contributions for water and methane on PaiNN and SO3Net models as a function of largest angle (θ) and opposed edge length (L) for triplets involved in a 3-body walk. | Nature Communications

Fig. 3: Average 3-body relevance contributions for water and methane on PaiNN and SO3Net models as a function of largest angle (θ) and opposed edge length (L) for triplets involved in a 3-body walk.

From: Peering inside the black box by learning the relevance of many-body functions in neural network potentials

Fig. 3

Panels a, c correspond to water and b, d to methane models. Panels a, b show the results for PaiNN-based, and panels c and d for SO3Net-based models. The relevance, in arbitrary units, is normalized by the absolute total relevance over the number of walks of the respective model and rescaled for each model type. A negative value implies a stabilizing interaction as the model output is the energy. Angle and distance ranges are divided in 50 bins each, and colors correspond to average values of the 3-body relevance for all triplets whose largest angle and opposed edge length correspond to the bin.

Back to article page