Sumner Harris1,Arpan Biswas1,Christopher Rouleau1,Alexander Puretzky1,Seok Joon Yun1,Gyula Eres1,Mina Yoon1,Rama Vasudevan1,David Geohegan1,Kai Xiao1
Oak Ridge National Laboratory1
Sumner Harris1,Arpan Biswas1,Christopher Rouleau1,Alexander Puretzky1,Seok Joon Yun1,Gyula Eres1,Mina Yoon1,Rama Vasudevan1,David Geohegan1,Kai Xiao1
Oak Ridge National Laboratory1
Synthesis of thin films has traditionally relied upon slow, sequential processes conducted with substantial human intervention, often using a mix of experience and serendipity to optimize material structure or discover properties. With recent advances in autonomous systems which combine automated synthesis, characterization, and artificial intelligence (AI), large parameter spaces can be explored autonomously at rates beyond what is possible by human experimentalists, promising to greatly accelerate our understanding of synthesis science. In this talk, I will discuss the infrastructure development of two highly versatile, autonomy-capable pulsed laser deposition (PLD) platforms: one based on real-time and <i>in situ</i> gas-phase and optical diagnostics and the other based around <i>in vacuo</i> robotic transfer for characterization. I will then show how we incorporated <i>in situ </i>and real-time diagnostics and characterization, a high-throughput methodology, and cloud connectivity to enable an autonomous synthesis experiment with PLD. We grew ultrathin WSe<sub>2</sub> films using co-ablation of two targets with real-time laser reflectivity for thickness control and showed a 10x increase in throughput over traditional PLD workflows. Bayesian optimization with Gaussian process regression was used with <i>in situ</i> Raman spectroscopy to drive synthesis and autonomously discover the growth window with a sparse sampling of only 0.25% of a broad 4D parameter space. Moreover, the process-property relationship predicted by the Gaussian process surrogate model suggests two different growth regimes and offers a more global understanding of the synthesis space than a human experimenter could gain without machine learning, information that is crucial for simultaneous, cooperative refinement of both synthesis experiments and theory. Any material that can be grown with PLD can be autonomously synthesized with our platforms and workflows, enabling AI-driven synthesis and accelerated discovery of a vast number of materials. This work was supported by the U.S. DOE, Office of Science, Materials Sciences and Engineering Division and was performed at the Center for Nanophase Materials Sciences, which is a DOE Office of Science User Facility.