Correctly selecting an equivalent circuit (EC) from electrochemical impedance spectroscopy (EIS) data is a key step in understanding charge transport, recombination, and diffusion in perovskite and related electronic materials. In practice, this task is often non-trivial because different physical processes can produce very similar impedance signatures, especially in the presence of experimental noise. In this work, we propose a machine learning-based approach for the automatic identification of 15 physically meaningful EC models covering a wide range of transport and diffusion behaviors. To train and benchmark the models, a large synthetic dataset consisting of 150,000 EIS spectra was constructed, where each spectrum was represented by 242 features derived from the real and imaginary impedance components across a broad frequency window, with frequency-dependent noise added to resemble experimental conditions.

Both deep learning and classical machine learning methods were explored. A one-dimensional convolutional neural network achieved an overall accuracy of 83.22% (macro F1-score of 0.83) and was able to capture local spectral trends, but its performance degraded for ECs with strongly overlapping Nyquist features, particularly those dominated by low-frequency diffusion. In contrast, tree-based models showed markedly better robustness. Among them, XGBoost delivered the highest accuracy of 95.24% with an F1-score of 0.95, followed closely by Random Forest with 93.80% accuracy. Detailed class-wise analysis indicated excellent recognition of circuits containing multiple time constants, while most errors arose from subtle overlaps between structurally similar semicircular arcs.

These results indicate that gradient-boosted tree models are especially well suited for high-dimensional EIS classification when physically informed features are used. The analysis also suggests that combining CNN-based feature extraction with XGBoost classification could further improve performance beyond 97%. To assess practical applicability, the framework was validated using experimentally measured EIS data from MAPbBr₃ single crystals grown by inverse temperature crystallization and measured under different applied biases. The XGBoost model consistently identified physically reasonable ECs, enabling stable fitting and reliable Mott-Schottky analysis, thereby confirming the effectiveness of the proposed ML-assisted strategy for experimental impedance studies.