| dc.description.abstract | Discovering new high-performance and functional materials is critical for the advancement of sustainable technologies. This thesis presents the design of a fully autonomous and self-driving laboratory for the discovery of optimized semiconductor materials used in solar photovoltaic applications. By leveraging sample miniaturization, parallelization, and machine learning-controlled robotics, the developed self-driving laboratory achieves a throughput of over 1,000 unique material compositions created, measured, and optimized per day.
The challenge of discovering a new optimal material comes down to the number of potential candidates within a search space. Perovskites — a class of semiconductors commonly studied for their properties as solar cells — have millions of possible compositional candidates within only an eight-dimensional search due to their widely interchangeable chemical formula. However, only a small fraction of these candidates have promise as high-performance semiconductors. While computational and predictive approaches provide a means to quickly reduce this search space, these methods alone often fail to accurately generalize to such vast and complex spaces due to few quality training datasets existing. Conversely, experimental approaches have high single-candidate accuracy but suffer from low throughputs. The presented self-driving laboratory fuses both computational and experimental approaches to rapidly down-select candidates with improved accuracy by iteratively generating its own experimental training dataset.
The developed self-driving laboratory is divided into four chapters: (1) synthesis, (2) characterization, (3) optimization, and (4) discovery. Each subsystem outlined in these chapters is engineered for speed and repeatability to maximize the cycles of learning when searching through vast and high-dimensional search spaces:
(1) Synthesis: A high-throughput synthesis tool is developed to automate the printing and crystallization of semiconductors by combining up to ten unique precursor formulations into target compositions, requested by a machine learning algorithm.
(2) Characterization: Two characterization systems are developed: one that rapidly parallelizes measurement of optical band gap properties using computer vision, and another that autonomously measures photoconductance electrical properties through contact-based measurement, controlled by neural network predictions.
(3) Optimization: A machine learning algorithm employing bounded search techniques is developed to guide the discovery of exceptional material compositions from high-dimensional and non-convex search spaces using the characterization results to suggest new optimal target compositions for synthesis in each subsequent iteration.
(4) Discovery: Lastly, control and multithreading communication protocols are developed to integrate all subsystems together with the goal of autonomously discovering optimal materials from a search space. Autonomous operation is achieved by coupling the subsystems with robotic arms and automated sample transporters, culminating in a fully self-driving laboratory, entitled DiSCO ([Di]scovery, [S]ynthesis, [C]haracterization, and [O]ptimization).
DiSCO creates, measures, and decides which new materials to make with full autonomy. A 24-hour optimization campaign is run to demonstrate the autonomy and performance of DiSCO with the objective of finding semiconductor compositions that have optimal optoelectronic properties for use as perovskite solar cells. The convergence to the optimization target as well as the number of candidate compositions synthesized from the search space are compared against experimentally derived results from literature. | |
