Active Learning for Tractable and Reproducible Pulsed Laser Deposition

TL;DR

An active learning framework based on Gaussian process Bayesian optimization that incorporates measured bulk and surface lattice properties along with impurity phase information to efficiently map the multidimensional growth space of LaVO$_3$ by PLD is used.

Abstract

This paper shows how data-driven machine learning approaches can improve growth control, reproducibility, and physical insight in the pulsed laser deposition (PLD) growth of correlated oxides. Despite well-known relationships between growth conditions and material properties, consistently producing high-quality films of complex materials like LaVO$_3$ remains difficult due to the highly non-equilibrium nature of PLD and the defects and competing phases that accumulate during growth. Here, we use an active learning framework based on Gaussian process Bayesian optimization that incorporates measured bulk and surface lattice properties along with impurity phase information to efficiently map the multidimensional growth space of LaVO$_3$ by PLD. By tuning the relative weighting of these properties, the model identifies an optimized region where phase-pure films of LaVO$_3$ exhibit two-dimensional surfaces, near-ideal lattice parameters, and minimal sub-band gap optical absorption. The trained model reveals clear competition among different defect formation mechanisms that are connected to unseen parameters like supersaturation and surface mobility, thus giving insight into the highly non-equilibrium process of PLD growth. Together, this demonstrates that property-guided machine learning can accelerate materials optimization while providing a new way to address fundamental growth mechanisms in PLD that enable understanding and utilization of quantum phenomena found in complex oxides.

Figures (15)

• Figure 1: Illustration of the experiment workflow: The center ellipse illustrates the structure of LVO and common defects, as labeled. The rest of the figure depicts the cyclic workflow of this study. The workflow starts with synthesis by PLD, which is followed by characterization. Then, the results are fed into an active learning model, which suggests growth parameters for the next iteration.
• Figure 2: (a) Sample XRD 2$\theta$-$\omega$ scans for several samples at different growth conditions. The dotted green vertical line indicates the XRD 2$\theta$ peak position that corresponds to the target $c$, the * symbol indicates the LVO (002) peak, the # symbol indicates the LaVO$_4$ (040) XRD peak, and the + symbol indicates the STO (002) peak. The labels (i-iv) indicate specific samples. (b) Rocking curve measurements of the LVO (002) peak of each sample. The solid red line is the best fit using a pseudo-Voigt function to quantify $\Delta \omega$. (c) AFM images of the surface morphology. The $R_{RMS}$ values are in units of nm.
• Figure 3: (a, b) Evolution of the mean (a) and acquisition function (b) of GP surrogate model showing the projection of the mean of the GP surrogate model onto the deposition temperature ($^\circ$C) and PO$_2$ (Torr) plane with increasingly more samples, as indicated at the top. White-to-red circles indicate sample scores. The red Xs indicate the maxima of the acquisition function after the last growth iteration in each column. The red X for growth iteration 29 is in the same position as the global optimum. The unfilled blue, purple and green squares correspond to the growth conditions of the samples given by the blue, purple and green curves in \ref{['fig:Figure_5_labels']}.
• Figure 4: (a-d) 2D projections onto the deposition temperature and PO$_2$ plane of the final GP surrogate from Fig. 3, but considering only a single type of measurement. Setting one of the weights ($\alpha$, $\beta$, $\gamma$, or $\delta$) equal to the value obtained by min-max scaling and the rest equal to zero, as indicated at the top of each panel, creates the plots. Plots of $\alpha$, $\beta$, $\gamma$, and $\delta$$\neq 0$ quantify film structure and defects, morphology, coherence of lattice planes, and the prevalence of the LaVO$_4$ phase, respectively. Lower GP mean values correspond to higher quality growth regions, and lower sample scores correspond to higher quality samples.
• Figure 5: Characteristics of film grown at global optimum: (a) Reciprocal space map about the (103) reflection in STO. The red triangle denotes the position of unstrained LVO. The dotted black line specifies the substrate peak position, and the dotted orange line indicates peak positions of unstrained or bulk lattice parameters. (b) XRD 2$\theta$-$\omega$ data. The dotted, vertical line indicates the 2$\theta$ position that corresponds to the target $c$. (c) Rocking curve measurements of the LVO (002) film peak on the same scale as \ref{['fig:Figure_2']} (b). (d) AFM measurement for the sample with $R_{RMS}$ in nm.