Skip to content

Improved prediction for failure time of multilayer ceramic capacitors (MLCCs): a physics-based machine learning approach

Summary

Multilayer ceramic capacitors (MLCCs) are ubiquitous in electronics; as dielectrics are made thinner and operating fields rise, reliability and mean time to failure (MTTF) estimation under accelerated stress becomes a central engineering problem. This APL Materials manuscript develops a physics-informed machine learning approach based on gradient-boosted trees (XGBoost) to predict MTTF for X7R MLCCs across temperature and voltage conditions, comparing against classical Eyring extrapolation and a newer tipping-point reliability model described in the dielectrics literature. A transfer-learning framing is used to improve accuracy when labeled HALT-style data are sparse for some stress splits and to support predictions for stress points without direct measurements. Penn State authorship includes Adri C. T. van Duin alongside MLCC reliability experts, linking the group’s multiscale materials footprint to electronic component lifetime modeling rather than atomistic reaction chemistry. The corpus PDF is an accepted manuscript; copy-edited version-of-record layout may differ.

Methods

Data-driven reliability modeling (not atomistic MD)

Supervised learning on HALT lifetime data; physics-informed features + XGBoost for MTTF vs Eyring/tipping-point baselines.

Transfer learning

Improves predictions when some T–V cells have sparse failure labels; targets extrapolation with uncertainty controls.

Materials context (continuum)

BaTiO\(_3\) MLCC processing motivates oxygen-vacancy degradation framing—ReaxFF is not the computational engine here.

1 — MD / ReaxFF (N/A). N/A for LAMMPS, NVT/NPT RMD, timestep values, E-field RMD, and metadynamics/umbrella—this work is HALT-data XGBoost reliability modeling (non-simulation “Methods” in the AGENTS review sense).

2 — ReaxFF training and static DFT (N/A). N/A as a ReaxFF parametrization or PBE+DFT kinetics study; see the APL Materials Methods/SI for feature design, train/test splits, and baselines (Eyring, tipping-point).

Findings

The abstract reports that the machine-learning model consistently outperforms both Eyring and tipping point models in the compared conditions, with favorable stability across splits. The approach is also described as yielding reasonable MTTF estimates for untested voltage–temperature combinations, supporting use as a practical reliability tool for X7R-class BaTiO₃ capacitors under accelerated stress. Claims about ranking and extrapolation should be read against the manuscript’s metrics sections and any SI tables in the published article. At a high level, the manuscript argues that physics-aware features help XGBoost respect known stress–life monotonicities and Arrhenius-like temperature trends while still learning nonlinear interactions between voltage, temperature, and manufacturing-driven variability that pure empirical laws miss. The framing is explicitly reliability-oriented: improve MTTF estimates where HALT data are abundant while still producing defensible predictions when some stress cells are sparse, which is common in industrial qualification campaigns. From a knowledge-base perspective, the paper is primarily a reliability ML contribution: it should not be read as atomistic BaTiO\(_3\) chemistry, even though oxygen vacancy narratives motivate degradation phenomenology at a continuum level.

Limitations

Accepted-manuscript PDF may differ from final publisher typesetting. The paper does not resolve atomistic degradation chemistry; it is a data-driven reliability model at component scale.

Relevance to group

Connects van Duin co-authorship to dielectric reliability and ML failure-time modeling for ferroelectric ceramic components.

Citations and evidence anchors