James Rondinelli, Northwestern University

"Towards active exploration of novel electronic materials"

Abstract: Over the last decade, functional electronic materials design has undergone a shift from chemical-intuition-based strategies to data-driven synthesis and simulation. Numerous machine learning (ML) models have been developed to successfully predict properties and generate new crystal structures. Many existing modalities, however, rely upon physical insights to construct handcrafted features (descriptors), which are not always readily available. For materials with sparse prior data and insufficient physical understanding, conventional ML models may display limited predictability. Here, I will address this challenge by introducing an adaptive optimization engine for materials composition optimization, where feature engineering is not explicitly required—so called featureless learning. I then describe a use case where multi-objective Bayesian optimization with latent-variable Gaussian processes is utilized to accelerate the design of electronic metal-insulator transition (MIT) compounds for memory applications [1]. I will then contrast this approach with supervised classification-based models for MIT compounds [2]. Last, I will highlight a recent quantitative study on structure-property relationship in crystal systems enabled by deep neural networks [3] and our efforts to mitigate bias in materials data using an entropy-targeted active learning (ET-AL) framework [4]. These findings enable the decoupling of structure and composition for future codesign of multifunctionality. Finally, I propose how integration of these different modalities could lead to harmonious iterative exploration of novel functional materials.

[1] Y. Wang et al. Appl. Phys. Rev. 7, 041403 (2020).
[2] A. Georgescu et al. Chem. Mater 33, 5591 (2021).
[3] Y. Wang et al. Phys. Rev. Research 4, 023029 (2022).
[4] H. Zhang et al. Appl. Phys. Rev. 10, 021403 (2023).