ML-Powered Impedance Analysis for Perovskite Single Crystals
Predicting low-frequency EIS response of halide perovskite single crystals - cutting 55-minute measurements to near-instant predictions using a supervised Random Forest model.
98.1%
Test accuracy (R² = 0.981)
8741
Experimental Datapoints
55 min
Measurement time eliminated per spectrum
4
ML models benchmarked
The problem
Electrochemical impedance spectroscopy (EIS) is the gold standard for characterizing perovskites — but recording a single spectrum down to 300 mHz takes up to 55 minutes. This prolonged exposure to light and bias causes the material to degrade during measurement itself, corrupting the very data being collected.
The low-frequency regime (below 1 kHz) is where crucial information about ion migration and carrier accumulation hides — but it's also the hardest and slowest part to measure.
Approach
Results
Prediction accuracy
The RF model achieved R² = 0.981 on the test set with RMSE of 0.0196 — accurately predicting both real (Z′) and imaginary (−Z″) parts of the impedance spectrum at unseen bias and illumination conditions.
Negative Capacitance Captured
The model correctly predicted the transition from capacitive to inductive behavior at high bias — a notoriously difficult feature that reflects ion dynamics. Errors were centered at zero for all three validation conditions.
Generalization to MAPbI₃
The model was tested on a completely different perovskite material (MAPbI₃). Error was one order of magnitude higher, demonstrating that material-specific ion dynamics are encoded in the data — and pointing toward future generalized models.
Time Saved
Recording a full EIS spectrum to 50 mHz takes ~55 minutes. The ML model generates the same low-frequency prediction in near-real time — while the material is protected from bias/light-induced degradation.