What question did this study set out to answer?

April 7, 2026Energy and AI0 citationsOpen Access

Accelerating Supercapacitor Lifetime Optimization Using Quantum Machine Learning Surrogates

MKMojtaba Khakpour Komarsofla AKAmirkianoosh Kiani

Key Points

The aim is to optimize supercapacitor lifetime prediction using quantum machine learning surrogates to reduce the reliance on extensive experimental data.
Developed a pipeline integrating cycling data with physical measurements and material characterization.
Constructed a unified dataset of over one million data points from fourteen supercapacitors.
Compared two surrogate models: classical kernel ridge regression and quantum-kernel ridge regression.
Employed Bayesian optimization to enhance prediction of cycle life.
Quantum-kernel ridge regression showed higher predictive accuracy in low-data scenarios.
With 75 training samples, Q-KRR achieved an R² of approximately 0.871, outperforming C-KRR’s 0.700.
Q-KRR-BO identified supercapacitors with an increased predicted cycle life, from 4609 to 6521 cycles.
Experimental validation indicated the optimized model achieved 10198 cycles at a 90% retention threshold.

Abstract

• Quantum ML surrogates improve supercapacitor cycle-life prediction • Quantum kernels outperform classical models in low-data regimes • Bayesian optimization accelerates lifetime design search • Unified electrochemical and materials dataset enables robust modeling • Experimental validation confirms optimized device performance Supercapacitor lifetime optimization is limited by the high cost and duration of long-cycle experiments, which constrains the amount of training data available for data-driven design. This study proposes and validates an experimental–computational pipeline that couples device-level cycling data with electrode physical measurements and material characterization, then uses kernel-based surrogate modeling and Bayesian optimization to maximize cycle life. A unified dataset was constructed by merging constant-current charge–discharge records with fixed physical/material descriptors, yielding more than one million raw data points across fourteen commercial supercapacitors tested under multiple operating protocols. Two surrogate models were evaluated under matched settings: classical kernel ridge regression (C-KRR) and quantum-kernel ridge regression (Q-KRR). In data-limited regimes, Q-KRR achieved higher predictive accuracy, with the clearest advantage at small training sizes (e.g., at n = 75, R 2 ≈ 0.871 for Q-KRR versus ≈ 0.700 for C-KRR). When embedded in Bayesian optimization, the quantum-driven pipeline (Q-KRR-BO) consistently proposed higher predicted cycle-life candidates than the classical pipeline (C-KRR-BO) under the same iteration budget (e.g., at n = 75, best predicted cycle life increased from 4609 to 6521 cycles). The optimized outcomes were experimentally validated using an independent device, which achieved 10198 cycles at the 90% retention threshold and was closer in feature space to Q-KRR-BO recommendations than to C-KRR-BO recommendations. Overall, the results indicate that quantum kernels can reduce the amount of required experimental training data for lifecycle optimization, while classical KRR remains preferable when runtime is the dominant constraint.

Perguntar à IA

Bookmark

View Full Paper