⚡ electrical engineering

A Simple, Modular Pipeline for Automated EIS Equivalent-Circuit Fitting via PCA- Space Nearest-Neighbour Matching, Validated Against Manual Fitting

This paper presents a modular, automated pipeline for EIS equivalent-circuit fitting that utilizes PCA-based nearest-neighbour matching against a synthetic database to generate robust starting parameters for multiple topologies, thereby avoiding premature model commitment while achieving higher accuracy and speed than manual fitting.

Original authors: Sanja Martinez, Lea Hrženjak

Published 2026-10-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sanja Martinez, Lea Hrženjak

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Electrochemical impedance spectroscopy is a way of listening to the electrical heartbeat of materials. By sending a tiny electrical signal through a substance and watching how it resists and shifts that signal across different speeds, scientists can learn about what is happening inside. It is like shining a light on a complex machine to see how its gears turn, revealing details about corrosion, battery health, or protective coatings without taking anything apart. However, turning these electrical signals back into a clear picture of the material's inner workings is notoriously difficult. The signals often look the same whether they come from one simple process or several complicated ones happening at once. For decades, experts have had to manually guess which mathematical model fits the data, a slow and subjective process that risks missing the true answer if the wrong guess is made first.

A team of researchers at the University of Zagreb has developed a new way to automate this difficult task, one that avoids the trap of guessing too early. Instead of forcing a single answer, their system keeps multiple possibilities open until the very end. They built a massive library of theoretical electrical signals, generated from just a few real-world examples that experts had already analyzed by hand. This library acts as a reference map. When a new, unknown signal arrives, the system does not try to solve the puzzle from scratch. Instead, it quickly scans its library to find the closest matches, using a mathematical technique that compresses the complex data into a simpler space to find neighbors. This step provides a shortlist of likely models and gives them a sensible starting point.

The real innovation lies in what happens next. Rather than accepting the first match, the system takes every model on that shortlist and runs a rigorous, independent test to see how well it actually fits the new data. It calculates a score for each attempt, measuring the tiny differences between the model and the real measurement. The system then presents all the surviving candidates side by side, showing the expert exactly which models fit best and how close they are to one another. This approach acknowledges that different electrical circuits can sometimes produce nearly identical signals, a problem that has long plagued the field. By refusing to commit to a single topology before checking the fit quality, the method tolerates this ambiguity and leaves the final decision to the human expert, armed with all the evidence.

The researchers tested this pipeline on a variety of signals, from those used to build the library to completely new measurements from different materials. In cases where the new signals fell within the range of the library, the automated system matched the results of careful human experts almost perfectly, agreeing on the values of the internal components within one percent. In fact, the automated fits were often more precise and mathematically tighter than the manual ones, taking only about six seconds to complete a task that usually requires much longer. The system proved robust enough to handle signals with overlapping electrical processes that are notoriously hard to separate, correctly identifying the underlying structure in most cases.

However, the study also clearly showed the limits of the approach. When the researchers tested a signal that was significantly different from anything in their library—specifically one with a much higher electrical resistance than the library contained—the system failed, producing a wrong answer. This was not a flaw in the logic but a boundary of the data. The researchers demonstrated that the system is modular and easy to fix: by adding just one new example of that difficult type of signal to the library, the system immediately learned to handle it correctly without needing to be rebuilt. This confirmed that the method works as long as the reference library covers the types of materials being studied.

The work offers a practical path forward for a field that has struggled with the non-uniqueness of its solutions. It does not claim to have solved the fundamental physics problem of distinguishing between identical signals, but it provides a reliable tool that surfaces all the plausible answers rather than hiding them behind a single guess. By combining a quick search with a thorough, independent verification of every candidate, the pipeline turns a subjective art into a transparent, reproducible process. It allows scientists to see the full landscape of possibilities, ensuring that the final interpretation of a material's electrical behavior is based on the best available evidence, not on the first guess that comes to mind.

Technical Summary: A Simple, Modular Pipeline for Automated EIS Equivalent-Circuit Fitting via PCA-Space Nearest-Neighbour Matching

Problem Statement
Automated interpretation of Electrochemical Impedance Spectroscopy (EIS) data faces two primary challenges: identifying the correct equivalent-circuit topology and supplying physically plausible starting parameters for nonlinear fitting routines. A significant hurdle is the non-uniqueness of fitting results, where distinct topologies can produce mathematically identical or near-identical spectra, or where parameters trade off against one another due to limited frequency ranges. Existing automated methods often commit to a single topology before evaluating fit quality or rely on supervised classifiers trained on synthetic data that fail to generalize to real, noisy experimental data from different instruments. Furthermore, many approaches do not adequately address the initialization problem for Complex Nonlinear Least Squares (CNLS) solvers, risking convergence to local minima.

Methodology
The authors propose a modular pipeline that avoids early commitment to a single topology. The workflow, implemented in the Orange data mining platform, consists of three main stages:

Synthetic Reference Database Construction:
- The process begins with a small number of manually fitted "seed" spectra (three measured spectra fitted under five different topological assumptions: 1TC, 2TC_SER, 2TC_PAR, 3TC_SER, 3TC_PAR).
- A synthetic database is generated analytically from these seeds by perturbing parameters within specific bounds (multiplicative for resistances/capacitances, additive for CPE exponents).
- This generation includes systematic single-parameter sweeps and randomized joint sampling, resulting in over 50,000 synthetic spectra.
- All spectra are projected into a shared Principal Component Analysis (PCA) space. The first five principal components (PC1–PC5) capture 97.62% of the variance.
Nearest-Neighbour Search and Candidate Generation:
- Any new EIS spectrum to be fitted undergoes the same normalization and PCA projection.
- A computationally inexpensive nearest-neighbour (NN) search identifies the 400 closest entries in the synthetic database.
- This step serves a dual purpose: it shortlists plausible candidate topologies based on which topologies appear in the nearest neighbours, and it supplies physically plausible starting parameters for each candidate topology derived from the matched neighbours.
Independent Refinement and Selection:
- Every topology represented in the shortlist is independently refined using CNLS, initialized by the parameters of the matched neighbours.
- The final selection is made post hoc based on the reduced chi-squared ( $\chi^2$ ) value. The pipeline retains all competing topologies that survive refinement, allowing human experts to view the full set of mathematically equivalent candidates rather than a single forced output.

Key Results
The pipeline was validated across three levels:

Internal Consistency: When manually fitted reference spectra were passed through the pipeline, the system correctly identified the exact reference circuit and recovered parameters within 0.4% to 0.9% of the manual fits.
Known Measured Spectra: For the actual measurements used to seed the database, the automated fits achieved lower reduced $\chi^2$ and lower parameter errors than the manual fits. The automated workflow reproduced manual results within 1% for most parameters.
Unknown and Genuinely Unknown Spectra:
- For "unknown" spectra (same instrument, different material), the pipeline correctly identified topologies and matched manual fits within 1% for all parameters, even for poorly constrained parameters (e.g., high uncertainty in $R_2$ ).
- For "genuinely unknown" spectra (different measurement systems), the pipeline successfully handled cases with non-identifiable parameters (where data only sets a lower bound) and near-degenerate time constants, converging to the same parameter values as manual fits.
- Limitations and Extensibility: The pipeline failed on a spectrum with impedance values outside the normalization envelope of the reference database. However, adding a single new seed spectrum representative of this regime to the database restored accurate fitting without altering the pipeline code. This demonstrated the system's modular extensibility.
- Computational Efficiency: The pipeline processes a spectrum in approximately 5–6 seconds (including NN search and CNLS refinement for multiple topologies).

Significance and Claims
The authors claim that this approach offers a practical solution to the non-uniqueness problem in EIS fitting by treating topology selection and parameter initialization as a unified retrieval problem rather than a sequential classification task. Key contributions include:

Tolerance of Topology Overlap: The method tolerates substantial overlap between structurally similar topologies (e.g., Series vs. Parallel 2TC) in the PCA space. A misleading nearest-neighbour match is not fatal; it merely results in a suboptimal initial guess that is discarded during the subsequent CNLS refinement if it does not yield the best fit.
Guidance over Replacement: Aligning with guidelines from Orazem and Ulgut, the tool is designed to guide the electrochemist by surfacing all competing candidates with their fit qualities, rather than dictating a single "best" model.
Modularity: The system is explicitly designed to be extended. When the reference database lacks coverage for a specific system (an "out-of-envelope" case), performance can be restored by adding new seed spectra, leaving the existing pipeline logic unchanged.
Robustness: The automated pipeline consistently achieved lower $\chi^2$ and parameter errors than manual fitting, suggesting that the NN-supplied initializations are superior to arbitrary or random starting points often used in manual workflows.

The paper concludes that while the pipeline cannot overcome fundamental physical limits (such as non-identifiable parameters due to frequency range limitations or near-degenerate time constants), it provides a robust, fast, and extensible framework for automated EIS analysis that respects the inherent non-uniqueness of the problem.

More like this