École Nationale Supérieure de
Techniques Avancées

Résumé

Materials or molecules formed by combining different chemical elements in various proportions and spatial arrangements hold valuable information for predicting atomistic properties. Yet, the vast number of possible structures poses significant challenges to efficiently exploring these datasets. To address this, the Chemiscope library was initially designed as a visualisation tool to aid in exploration of these datasets. With the new development presented in this rapport, we streamline the process, enabling to explore data without additional preparation steps for representation and dimentionality reduction of the features. We also introduce multiple improvements of functionalities, specifically, the ability to switch between properties targeting the local environment or entire structure. Additionally, we implemented the possibility to embed interactive Chemiscope widgets into documentation automatically generated by Sphinx-Gallery by developing a specialised extension. Furthermore, we have compared the mean absolute error and the inference time by training the Point Edge Transformer models on the MC3DL dataset with variations in the hyperparameters related to the transformer model dimension and the number of GNN layers.