Hybrid Organic-Inorganic Perovskites (HOIPs) have been widely researched over the last
decade, with great success in increasing their solar cell efficiency. HOIPs offer a viable
alternative to silicon solar cells, primarily in their promise for large-scale fabrication via solution
processing. Unfortunately, too much remains a mystery regarding a fundamental
understanding of the self-assembly process and precisely which solution processing variables
would be ideal (reagents, solvents, and processing conditions). This makes it a challenge to
create a member of the HOIP class with pre-specified characteristics beyond simply maximizing
device efficiency. There exists an overwhelming combinatorial problem due to the variety of
possible species: Many choices of A and B site cations (including blends of each), the choice of
halide (again, potentially a blend), choice of solvent blend, and aspects of their processing
(timing of adding anti-solvents, etc.). Exploring all possible candidates would be insurmountable
using a standard Edisonian approach. To redress this situation, we have developed a model
tailored to be used within a Bayesian optimization Gaussian Process Regression (GPR) scheme.[1]
Bayesian optimization is, arguably, the best method of maximizing an expensive objective, like
ours, and is ideal for this scenario. We tested this approach by maximizing the solvation of
HOIP salts (the smallest APbX3 molecule) using a simple computational ersatz for the enthalpy
of solvation, namely the intermolecular binding energy between the salt and solvent obtained
from DFT calculations, which served as our objective function. We hypothesized a connection
between good solvation and slow crystallization kinetics, and hence high quality thin films. We
studied two test cases for which we were able to compute the DFT intermolecular binding
energy of all possible candidates. The first was the best species composition for a pure halide in
a pure solvent (72 possible combinations). The second was for a mixed halide in a pure solvent
(240 possible combinations). We found we were able to achieve global maximization in under
half the number of iterations needed for the next best model (a Bayesian optimization
approach without using our model), and significantly better than all other common models.
We were able to identify the experimentally preferred “gold standard” compositions for pure
halides and mixed halide systems in a variety of pure solvents. Further, we were also able to
identify unexpected highly ranked predictions of high performers that deserve experimental
confirmation.