EXTENDED ABSTRACT: The foundation of materials science is encapsulated in a multifaceted tetrahedron comprising: "composition-structure-processing-performance" four corners. Material performance is influenced by the interplay of complex physical mechanisms, promoting interdisciplinary research involving materials science, physics and AI algorithms. The success of AI for materials science is significantly influenced by the quality of experimental data, the selection of material descriptors, and AI algorithms utilized. Notably, the 2024 Nobel Prize in Physics was awarded to the pioneers in this field. It has been realized in the field that existing neural network algorithms perform inadequately with small datasets and high-dimensional descriptors. This issue arises because limited data makes it difficult to effectively train numerous hyper-parameters. To overcome the problem, we employed a novel Hierarchical Neural Network (NNH) to performance statistical ensemble integration for a large number of neural network sub-models with reduced descriptor combinations. Each sub-model at the optimal dimension for a given dataset retains unique correlations between material descriptors and target property and can be trained in parallel by the same dataset with limited data points. HNN algorithms then extract all useful information from millions of neural network sub-models. This approach solved the contradiction between small dataset and high dimensional relevant descriptors. We will discuss the applications of this superconductivity, catalytic reactions, amorphous formations and other fields.
Keywords: High dimension of descriptors, small data sets; NNH model; statistical ensemble integration