UC Berkeley

Synthesis is the major bottleneck in the computational design of new inorganic materials for technological applications, including rechargeable batteries, catalysis, sensing, photovoltaics, and more. Unlike molecular reactions, where individual bond-breaking or formation steps can be explicitly followed, reactions involving solid inorganic materials happen at great spatial scales where atomic control cannot be achieved. This is especially true for solid-state synthesis, which involves the reaction of bulk solid powder precursors at elevated temperatures (i.e., the ceramic method). Solid-state synthesis is an especially attractive approach given its simplicity and scalability to bulk chemical manufacturing. Recent work has focused on the experimental characterization of reaction pathways in solid-state synthesis to elucidate theoretical principles determining the performance of the synthesis reactions. These works have shown that powder reactions often involve a significant number of intermediates, including crystalline phases that sometimes persist as undesirable impurity products. Meanwhile, databases of crystalline materials and their properties have rapidly grown with the development of high-throughput density functional theory calculation workflows. These databases enable the study of phase equilibria, reaction energetics, and predicted products in solid-state reactions.

In this dissertation, we address the current deficit of theoretical models for predicting reaction pathways in solid-state synthesis, taking advantage of the aforementioned breadth of computed thermodynamic data. We focus on the development of both forward and inverse prediction models, with the goal of identifying synthesis reactions that selectively produce desired target phases with minimal competing intermediates. The central contribution of this work is the development of a reaction network model and methodology for predicting reaction pathways and key intermediate phases encountered in solid-state synthesis. The solid-state reaction network is a large graph-based network created from a mixture of computed and experimental free energies, supplemented by a machine-learned descriptor for Gibbs free energy. Using standard graph pathfinding approaches, we perform forward prediction, suggesting a ranked set of likely reaction pathways in the synthesis of four inorganic materials (YMnO3, Y2Mn2O7, Fe2SiS4, and YBa2Cu3O6.5). We compare the network-predicted pathways to experimental results characterized in the literature and find that the reaction network model captures key intermediates and reproduces the experimental reaction pathways for several targets.

We then address the current challenge of predicting the selectivity of solid-state reactions. In a detailed study of the synthesis of the Y2Mn2O7 pyrochlore, we show that assisted metathesis reactions with precursors 3 A2CO3 + 2 YCl3 + Mn2O3 selectively produce Y2Mn2O7, but only when sodium is used as the alkali metal (A=Na). We explain the unique selectivity of the Na-based reaction from observation of the formation of an NaxMnO2 intermediate with in situ synchrotron X-ray diffraction. It is the stability of this intermediate into high oxygen chemical potential ranges that facilitates direct reaction with YOCl intermediate to produce Y2Mn2O7. Using this observation, we develop a selectivity metric related to the change of chemical potentials of the precursors that rationalizes the unique selectivity of the NaxMnO2 intermediate due to its direct connection in chemical potential space with Y2Mn2O7. We improve upon this idea further and suggest two reaction selectivity metrics (primary and secondary competition) measuring the degree of competitiveness of impurity reactions acquired from the solid-state reaction network. We validate the use of these metrics by applying them to reactions previously documented in the literature, as well as for planning our own experiments synthesizing BaTiO3 from eleven different precursor combinations spanning a range of chemistries, free energies, and selectivities. Our work culminates in the development of an inorganic synthesis planning workflow that can be used to suggest precursors that optimally yield a desired target.

Along the way, we devise a possible explanation for related experiments that produce two metastable polymorphs: orthorhombic YMnO3 and layered α-NaFeO2. Our hypothesis centers around the fact that metastable polymorphs may be stabilized in certain environmental conditions due to differences in defect accommodation between the nominally stable and metastable polymorphs. For both YMnO3 and NaFeO2, we show, with charged point defect density functional theory calculations, that the metastable polymorphs can be targeted in oxygen-rich environments. Altogether, the application of theoretical concepts and predictive models developed herein will greatly enhance both human-led and machine-driven synthesis exploration efforts, reducing the synthesis bottleneck.