Autonomous Materials Exploration by Integrating Automated Phase Identification and AI-Assisted Human Reasoning

TL;DR

An autonomous materials synthesis extension to SARA, the Scientific Autonomous Reasoning Agent, is presented, utilizing phase information provided by an automated probabilistic phase labeling algorithm to expedite the search for targeted phase regions.

Abstract

Autonomous experimentation holds the potential to accelerate materials development by combining artificial intelligence (AI) with modular robotic platforms to explore extensive combinatorial chemical and processing spaces. Such self-driving laboratories can not only increase the throughput of repetitive experiments, but also incorporate human domain expertise to drive the search towards user-defined objectives, including improved materials performance metrics. We present an autonomous materials synthesis extension to SARA, the Scientific Autonomous Reasoning Agent, utilizing phase information provided by an automated probabilistic phase labeling algorithm to expedite the search for targeted phase regions. By incorporating human input into an expanded SARA-H (SARA with human-in-the-loop) framework, we enhance the efficiency of the underlying reasoning process. Using synthetic benchmarks, we demonstrate the efficiency of our AI implementation and show that the human input can contribute to significant improvement in sampling efficiency. We conduct experimental active learning campaigns using robotic processing of thin-film samples of several oxide material systems, including Bi$_2$O$_3$, SnO$_x$, and Bi-Ti-O, using lateral-gradient laser spike annealing to synthesize and kinetically trap metastable phases. We showcase the utility of human-in-the-loop autonomous experimentation for the Bi-Ti-O system, where we identify extensive processing domains that stabilize $δ$-Bi$_2$O$_3$ and Bi$_2$Ti$_2$O$_7$, explore dwell-dependent ternary oxide phase behavior, and provide evidence confirming predictions that cationic substitutional doping of TiO$_2$ with Bi inhibits the unfavorable transformation of the metastable anatase to the ground-state rutile phase. The autonomous methods we have developed enable the discovery of new materials and new understanding of materials synthesis and properties.

Figures (15)

• Figure 1: The overall workflow of SARA-H. (a) The autonomous experimental workflow or SARA-H. The leftmost panel represents SARA-H's synthesis technique, which is based on the annealing of amorphous thin film samples with lg-LSA under specific conditions. A 3D-rendered experimental setup is shown, with the laser to the left, a camera to the right representing characterization and the power profile in the backdrop, and the thin-film sample mounted on a stage. An annealed lg-LSA stripe (second panel) is then characterized with narrow-beam synchrotron X-ray diffraction. The top image shows a micrograph of an annealed stripe, together with the nominal XRD footprint that is scanned across the stripe, while the bottom image presents the corresponding XRD heat map. The third panel illustrates analysis of the diffraction data on-the-fly using the probabilistic muiltiphase-labeling algorithm CrystalShift. Finally, the expected phase fraction from the probabilistic phase model can be utilized to construct a phase diagram (shown as a probability distribution in a high dimensional processing space), which is progressively improved using Bayesian optimization. The AL loop is closed by performing the next, optimal synthesis. (b) The optimization strategy may be augmented with human expertise to further enhance the targeted synthesis of specific phases. For a given material system, the expert prepares a list of candidate phases to study in the active learning run. During the run, the expert monitors results and determines which candidate phases to retain in the the active learning loop for optimization. This human-in-the-loop behavior creates an interface for experts to control the behavior of the autonomous experimentation. Details on the time scales of every step involved in the SARA-H workflow are compiled in Sec. \ref{['sec:timescales']} of the Supplemental Materials.
• Figure 2: Performance of various active learning sampling strategies in a synthetic phase space. (a) The median regret on one of the phases for a materials system containing a total of 8 phases against the number of iterations with different sampling strategies. Solid lines are the stripe variant while the dashed lines represent the single-point counterpart. (b) The IQR of the regret versus the number of iterations for each sampling strategy, for the same phase as in (a). Compared to random sampling, the regret IQR for other methods rises and decreases earlier and more sharply. This indicates the efficient exploration and exploitation of active learning methods. (c) The enhancement factors represent the ratio between the objective values obtained from AL and those achieved through random sampling at various stages, averaged over all 8 phases present in the materials system. This quantity illustrates the degree to which the results improve when utilizing AL techniques. (d) The acceleration factors represent the ratios of the average iterations needed to achieve the same objective value for random sampling with respect to each AL strategy, averaged over all 8 phases that exist.
• Figure 3: (a) Processing phase diagrams of the Bi2O3 system after SARA-H's converged AL cycles with 95 iterations. Blue and purple domains represent the presence of $\beta$-Bi2O3 and $\delta$-Bi2O3, respectively. The saturation of each region is proportional to the corresponding phase activation with amorphous domain represented as uncolored regions. The dashed black lines approximate the phase boundaries and are intended as a guide for the eye. (b) $R_s^2$ convergence as a function of the AL iterations for each of the two observed phases, with their average shown as the dashed line.
• Figure 4: (a) Gibbs triangle of the ternary phase space Bi-Ti-O. The diagram features all compositions considered in the candidate pool for our probabilistic phase determination algorithm, represented as either circles or crosses. Circles indicate compositions that were observed in the experiment while crosses denote phases that did not form. Note that most candidate compositions lie along the tie line between Bi2O3 and TiO2, thereby ensuring appropriate charge balance. (b) Equilibrium phase diagram based on Ref. kargin_phase_2015. The compositional axis is rescaled from the original figure in Ref. kargin_phase_2015 to match the atomic fraction convention employed in this manuscript.
• Figure 5: Stages of active learning for the Bi-Ti-O autonomous experimentation.