Supervised learning applied to high-dimensional millimeter wave transient absorption data for age prediction of perovskite thin-film

TL;DR

This paper demonstrates that supervised machine learning, specifically Gaussian Process Regression with a Matern kernel, can accurately predict the age of perovskite thin films from millimeter wave transient absorption data, revealing aging-related parameter shifts.

Contribution

It introduces a novel application of supervised learning to high-dimensional millimeter wave data for aging prediction of perovskite films, identifying key parameters and optimizing the GPR model.

Findings

GPR model achieves R^2 around 0.97 for age prediction.

Five parameters strongly correlate with sample age.

The Matern-5/2 kernel provides the best predictive performance.

Abstract

We have analyzed a limited sample set of 120 GHz, and 150 GHz time-resolved millimeter wave (mmW) photoconductive decay (mmPCD) signals of 300 nm thick air-stable encapsulated perovskite film (methyl-ammonium lead halide) excited using a pulsed 532-nm laser with fluence 10.6 micro-Joules per cm-2. We correlated 12 parameters derived directly from acquired mmPCD kinetic-trace data and its step-response, each with the sample-age based on the date of the experiment. Five parameters with a high negative correlation with sample age were finally selected as predictors in the Gaussian Process Regression (GPR) machine learning model for prediction of the age of the sample. The effects of aging (between 0 and 40,000 hours after film production) are quantified mainly in terms of a shift in peak voltage, the response ratio (conductance parameter), loss-compensated transmission coefficient, and the radiofrequency (RF) area of the transient itself (flux). Changes in the other step-response parameters and the decay length of the aging transients are also shown. The GPR model is found to work well for a forward prediction of the age of the sample using this method. It is noted that the Matern-5 over 2 GPR kernel for supervised learning provides the best realistic solution for age prediction with R squared around 0.97.