Page 46 - Read Online

Page 6 of 15 Allen et al. J Mater Inf 2024;4:35 https://dx.doi.org/10.20517/jmi.2024.72

inferred by comparing their characteristic length scales. A smaller length scale indicates that the model is
more sensitive to changes in that feature, meaning that small variations in the feature value lead to
significant changes in the model’s output. Conversely, a larger length scale suggests that the feature varies
more slowly and has a less pronounced impact on the model’s predictions. Supplementary Table 3 shows
that the length scale hyperparameters of pulse length and radiant energy are ~2 to 3× shorter than those of
the number of µpulses and duty cycle, indicating that pulse length and radiant energy are more important.

The noise variance for each GP model was calculated as the standard deviation for each similarity metric
from the LHS PC condition of which most samples were produced. We consider this value to be the
uncertainty in making reproducible PC MAPbI films. Noise variance values were held constant during
3
active learning iterations as the measurement uncertainty was not expected to change with the addition of
new samples to the dataset. The acquisition function for the models was the upper confidence bound
(UCB). Following the literature [12,33,34] , a UCB exploration hyperparameter of b = 1 was used to maintain a
balance between exploration and exploitation when picking the next condition. Each similarity metric
model would suggest a new condition to try; thus, in each BO iteration, we have a total of four new
conditions, one from each GP model. The search for optimized PC conditions is declared successful when a
PC condition produces similarity metric values comparable to the values for two TA MAPbI films (~10
-10
3
-2
-1
Procrustes distances, < 2.0 × 10 Fréchet distance, and < 2.0 × 10 RMSD); i.e., the similarity is within the
experimental uncertainty.
Grain size determination
In literature, the ASTM E112-13 line intercept method is the standard for determining average grain size
[21]
for crystalline samples. However, the ASTM method is a cumbersome process that often requires manually
placing several random line segments onto an image and counting the number of grain boundaries that are
crossed. The average for a single image can then be calculated after tabulating the total length of the line
segments and the total number of boundaries crossed. The drawbacks of this method to analyze multiple
images include its time-consuming and tedious nature, the use of only limited data, and possible bias from
the researchers in choosing the lines. In this study, we propose an alternative approach whereby a set of
images of the same size and magnification can be analyzed in minutes. Using an artificial intelligence (AI)
segmentation model derived from Facebook’s open-source Segment Anything Model in a Google
[35]
Collaborate environment, we can generate masks that correspond to the location of crystalline grains within
an SEM image. We can then extract the size of the grains and quickly display the information within our
script. The results from the AI segmentation model are compared to those from the ASTM E112-13
method. Our AI segmentation method is much faster in processing multiple SEM images and uses all data
in the image.

RESULTS AND DISCUSSION
UV-vis spectra
Supplementary Table 1 displays the input variables of the PC conditions and the four similarity metrics
when compared to the TA sample made at the same time. All PC conditions mentioned in the rest of this
study will be referenced as PC ## where “##” represents the number of the PC condition as labeled in
Supplementary Table 1. The first 20 rows (PC 00 to PC 19) constitute the LHS conditions used as the
training dataset for the initial GPR models. Of note from the LHS conditions was PC 07. Condition PC 07
(13.2 J/cm ) was a particularly high radiant energy pulse that ablated the MAPbI from the substrate upon
2
3
exposure and suggested an upper limit to the allowable radiant energy delivered to the film.