UC Berkeley

Advanced inorganic materials are frequently designed as solid-solutions, in which elemental substitution within a host crystal framework tunes functional properties such as ionic conductivity. Exploring these composition spaces by solid-state synthesis is constrained by phase competition, and while Density Functional Theory (DFT) can screen thousands of candidates in hours, realizing them experimentally remains manual and slow. This dissertation describes software and hardware systems that couple DFT-derived priors to automated experimentation, and that feed experimental outcomes back to refine those priors. Chapter 2 introduces AlabOS, a Python-based workflow-management framework for autonomous laboratories. AlabOS separates workflow definitions from device drivers through four abstractions (samples, devices, tasks, experiments) and enforces resource safety across concurrent operations through a dynamic reservation system. A simulation mode runs the same workflows with mocked drivers, allowing logic and race conditions to be tested before deployment. AlabOS has run continuously since 2022 and has orchestrated more than 3,500 samples across multiple platforms. Reliable orchestration, however, is only useful to the extent that the laboratory can also carry out its physical steps, and the most stubborn of these is the manipulation of dry post-heated solids. Chapter 3 describes the Dry Automated Solid Handler (DASH), a modular robotic station that bridges high-temperature furnaces and powder X-ray diffraction (XRD). DASH performs automated solid retrieval, solvent-free grinding, targeted sieving, and dynamic flattening of post-heated products. Across 1,588 samples spanning 862 chemical spaces, each loaded with up to 1.8 g of starting material (more than 94% above 1 g), DASH recovered at least 300 mg of material from 66% of samples, establishing a lower-bound operational success rate, with an average mass recovery of 43% of the estimated starting mass. On Li2MnO3 and Li2TiO3 benchmarks, DASH-prepared specimens yield XRD signal-to-noise ratios and peak widths comparable to expert-prepared samples, with minor residual variation in peak broadening. With orchestration and solid handling in place, the laboratory can be driven by computational design across a range of chemistries. Chapter 4 presents the A-Lab, an autonomous laboratory that combines AlabOS, DASH, DFT-derived priors from the Materials Project and the GNoME dataset, text-mined historical recipes, an active-learning recipe planner, and automated phase analysis. Over 17 days of continuous operation, the A-Lab attempted 57 target compositions through 353 synthesis recipes, realized 36 of the 57 targets, and identified four dominant failure modes in the remaining 21: precursor volatility, sluggish reaction kinetics, amorphous phase competition, and residual DFT inaccuracies. The A-Lab is well suited to broad screening over discrete targets, but systematic navigation of continuous solid-solution domains, where a specific property must be optimized while respecting synthetic accessibility, requires a more targeted strategy. Chapter 5 introduces the Cost-guided Autonomous Solid-state Synthesis (CASS) framework for continuous solid-solution spaces. CASS combines DFT-derived thermodynamic priors, structural compatibility rules, and ionic-conductivity surrogates into aggregated cost functions that are updated after each synthesis from Rietveld-refined phase identification. Deployed in the A-Lab for multicomponent Na superionic conductor (NASICON) electrolytes, CASS ran 78 trials over four days, of which 23 yielded NASICON as the major phase and 18 fell within the high-conductivity target region. Two representative compositions, a Hf-Zr-In-Ga phosphosilicate and a Hf-In-Ga phosphosilicate, showed bulk ionic conductivities of 3.96 and 3.17 mS/cm, respectively. Together, these chapters address four bottlenecks in autonomous inorganic synthesis: workflow orchestration, dry powder handling, closed-loop integration of prediction and experiment, and navigation of continuous solid-solution spaces. The resulting infrastructure supports both broad screening of diverse synthesis targets and focused optimization of functional compositions in continuous design spaces, and it returns structured experimental feedback, including failure modes and empirical metastability limits, that can be used to refine future computational priors.