Fig. 1. The frameworks of Electron Configuration models with Stacked Generalization (ECSG) and Electron Configuration Convolutional Neural Network (ECCNN).

a ECSG yields predictions of thermodynamic stability by integrating predictions from three complementary base-level models into a meta-level model. The base-level models include: (i) ECCNN uses convolutional neural networks to extract complex features related to the electron configuration; (ii) Magpie emphasizes the importance of including statistical features derived from various atom properties, such as atomic number, atomic mass, and atomic radius³⁰; (iii) Roost conceptualizes the chemical formula as a complete graph of elements, employing graph neural networks to learn the relationships and message-passing processes among atoms⁴⁸. b The architecture of ECCNN. It begins with input features arranged in a 3D matrix, followed by two convolutional layers. Then, a flatten layer converts features extracted from convolutional layers into a 1D vector. Finally, a multilayer perceptron (MLP) processes the flattened features through fully connected layers to output the final predictions.