Automating Battery Discovery through Data Infrastructure and Scientific Workflows

The urgent need to discover new materials to reduce CO₂ emissions and toxicity, while accounting for resource scarcity and geographical distribution, motivates accelerated research strategies. Ab initio computation and Artificial Intelligence (AI), leveraging modern computational power, create new avenues for acceleration. Embedded within Material Acceleration Platforms (MAPs), these tools together with experimental feedback enable closed-loop optimization that expedites discovery.

This thesis advances MAPs and scientific workflows for batteries through six contributions in the form of five studies and a forward-looking perspective. First, it outlines battery design and core components, along with contemporary material classes and their properties, and presents a review of solid-state inorganic electrolytes for potassium-ion batteries. The review highlights the current scarcity of computational studies and identifies potentially promising structural families and knowledge gaps.

Second, it addresses the concept of MAPs, in which methods such as AI, automated experimentation, and theoretical calculations are integrated into a single platform to accelerate the discovery and development of advanced materials. It presents the development and deployment of FINALES, a modular, problemagnostic Material Acceleration Platform (MAP) framework, and its use in an optimization campaign on electrolyte mixing ratios for EC, EMC, and LiPF6 formulations. The campaign spans scales, optimizing ionic conductivity at the material level and end-of-life performance of coin cells at the device level. Building on a presented overview of ontologies and their role in data management, the study semantically annotates and embeds the campaign’s data in a knowledge graph to support interoperability, provenance, and reuse.

Next, it offers a perspective on the state of MAP-based research and the maturity of core components for integration. Ontology-driven data infrastructures are found to be underdeveloped, while Self-Driving labs (SDLs) and AI agents are at intermediate maturity. The work argues that dynamic workflow managers are essential to advance both. As a demonstrator, the study integrates the PerQueue dynamic workflow manager into a small, modular, self-driving color-mixing platform.

Then, to further accelerate discovery, it explores high-throughput Density Functional Theory (DFT) workflows and their coupling to mesoscale transport models. While DFT enables rapid screening, bridging to experimental time and length scales remains challenging. As an initial step toward addressing this, it presents an ongoing work integrating kinetic Monte Carlo (kMC) simulations into an automated workflow for assessing transport properties. This is applied to Na and K Prussian blue analogue (PBA) with Ni, Fe, Cr, and Mn substitutions. Preliminary results indicate that Ni substitution at the Ni-coordinated site in the K Prussian blue analogues (PBAs) generally yields the lowest migration barriers among those tested, whereas the Cr-substituted system exhibits particularly high barriers. Intermediate stages in the workflow suggest that additional sampling is required for the K PBAs systems to enable reliable kMC simulations. In contrast, the preliminary Na PBAs results point to the need for broader model refinements.

Finally, it addresses interoperability and reproducibility in high-throughput ab initio workflows. Using four workflow managers and four DFT engines, the work implements five variants of the same open-circuit voltage (OCV) intercalation workflow. A universal I/O schema enables engine-agnostic execution and facilitates systematic analysis of DFT code-dependent discrepancies. While total energies for pristine structures align well across implementations, vacancy-containing structures introduce significant deviations. The study provides guiding principles for robust DFT workflow implementations in high-throughput settings and emphasizes how cross-code validation can enable corrective strategies.

The thesis concludes with a forward-looking perspective on key technologies to advance, pitfalls to avoid, and community practices to adopt, with an emphasis on standards, interoperability, and closer alignment between modeling and experiment.