Insights into the Effect of Bioreactor pH on Chinese Hamster Ovary Cell Metabolism and Site-Specific N-Linked Glycosylation VIA Experimental and Computational Approach

Multicomponent perovskite oxides are an essential class of inorganic materials with diverse applications, where the ordering of cations can profoundly influence their physical and chemical properties. Unfortunately, the discovery and optimization of such oxides have been significantly hindered by their compositional and structural complexity. State-of-the-art brute-force experimental or computational methods can hardly resolve such complexity efficiently. Even one multicomponent oxide composition can have tens to hundreds of cation orderings, which are too expensive for brute-force high-throughput screening.

To boost the high-throughput screening of multicomponent perovskite oxides, graph neural network models can be used to effectively predict material properties from the connections of atoms in crystal structures, allowing efficient computational discovery over a vast search space. However, while these models have been demonstrated to capture the compositional dependence of material properties across a broad chemical space, it is unclear whether such models can also effectively learn the ordering dependence of similar properties for a given composition. In addition, structures after costly density functional theory (DFT) relaxations have been dominantly used for training and leveraging such graph neural network models, which cannot truly eliminate the prohibitive cost of DFT in high-throughput screening.

In this work, we developed graph neural network models to accurately predict key cation ordering−dependent properties of multicomponent perovskite oxides from DFT-unrelaxed cubic perovskite structures. Essentially, models that infer material properties by indirectly learning the thermodynamics of mixing lead to lower errors than those that predict these properties directly. Moreover, equivariant neural networks [1,2] better capture the ordering dependence of such properties than their symmetry-invariant counterparts [3] due to their higher expressivity in distinguishing the difference between the coordination environments and symmetries of various orderings. Lastly, using contrastive losses in the training of these models further increases the prediction accuracy by facilitating an optimal balance between capturing both the compositional and structural dependence of material properties. Overall, this work highlights an effective strategy for designing machine learning models with optimized robustness and generalizability for discovering multicomponent materials.

References:

[1] Geiger, et al. arXiv:2207.09453 (2022).

[2] Schütt, et al. arXiv:2102.03150 (2021).

[3] Xie, et al. Phys. Rev. Lett. 120, 145301 (2018).