PoFINN: Integrating Physics of Failure and Neural Networks for Long-Term Degradation Forecasting of Aluminum Electrolytic Capacitors

Aluminum electrolytic capacitors play a critical role in power electronics, where degradation under storage conditions poses reliability challenges for long-lifecycle systems in defense, avionics, and industrial applications. This research introduces the physics-of-failure-informed neural network (<sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">PoFINN</small>), which integrates electrolyte evaporation physics with neural architectures for accurate degradation forecasting. Existing approaches struggle with limited data, nonstationary patterns, and physically implausible predictions. <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">PoFINN</small> addresses these limitations through two key innovations: an embedded first-principles evaporation model capturing capacitance loss through mass balance equations and a dual-phase training methodology. This approach initializes long short-term memory weights using simulated physical data, and then refines parameters with empirical measurements. A parametric hybrid loss function balances physical constraints with data errors, maintaining plausibility while optimizing accuracy. Experimental validation shows <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">PoFINN</small> outperforms conventional forecasting approaches, achieving error reductions of 27%–81% across all time horizons compared to statistical, neural, and ensemble techniques. The most substantial improvements (70%–80%) are observed against recurrent architectures, with advantages more pronounced in long-term forecasting. Performance remains robust across different train–test splits (50–50, 60–40, and 70–30), highlighting <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">PoFINN</small>’s effectiveness with limited data—addressing a critical challenge in degradation prediction. The methodology shows immediate practical value for inventory management in reliability-critical sectors and potential applicability to other components, improving maintenance planning and replacement decisions.