Matini-Net: Versatile\nMaterial Informatics Research\nFramework for Feature Engineering and Deep Neural Network Design

In this study, we introduced Matini-Net, which is a versatile\nframework\nfor feature engineering and automated architecture design for materials\ninformatics research using deep neural networks. Matini-Net provides\nthe flexibility to design feature-based, graph-based, and combinations\nof these models, accommodating both single- and multimodal model architectures.\nFor validation, we performed a performance evaluation on the MatBench\nbenchmarking dataset of five properties, targeting five types of regression\narchitectures that can be designed using Matini-Net. When applied\nto each of the five material property datasets, the best model performance\nfor the various architectures exhibited <i>R</i><sup>2</sup> > 0.84. This highlights the usefulness and flexibility of Matini-Net\nfor accelerating materials discovery. Specifically, this framework\nwas developed for researchers with limited experience in deep learning\nto easily apply it to research through automated feature engineering,\nhyperparameter tuning, and network construction. Moreover, Matini-Net\nimproves the model interpretability by performing an importance analysis\nof the selected features. We believe that by employing Matini-Net,\nmachine and deep learning can be applied more easily and effectively\nin various types of materials research.