Time-dependent dielectric breakdown (TDDB) is an important failure mechanism that affects semiconductor device reliability. The degradation lifetime depends strongly on both electric field and temperature, yet these dependencies are commonly treated using separate empirical acceleration factors. This work develops a thermodynamic description in which electrical stress changes the accessible microscopic states of the dielectric and therefore contributes an entropic term to the Gibbs free energy. The resulting formulation separates the field-independent activation barrier, $E_a$, from a field-dependent entropic contribution characterized by $\alpha$. This relation predicts a linear decrease in apparent activation energy with the logarithm of the electric field. Published TDDB data for SiO₂, high-k, low-k, and Cu/low-k dielectric systems exhibit this behavior. Independent temperature- and field-dependent measurements from the same source yield a common entropic field-energy scale of approximately 1.2 eV. A one-second reference field, $E_{1sec}$, provides a physical time anchor that allows field and temperature acceleration to be compared without assuming a common pre-exponential factor across temperatures. For an absolute time temperature series, independently extracted values of $\alpha$ vary by 11.9% across three temperatures, while a fixed α produces a normalized relation with R² ≈ 0.97. The parameters $E_a$, $\alpha$, and $E_{1sec}$ provide a compact thermodynamic representation of field- and temperature-dependent TDDB kinetics across the datasets examined here.

Power-law exponents are widely used to describe time-dependent transistor degradation and to extrapolate accelerated measurements toward operating conditions. Their physical interpretation, however, depends on the baseline from which degradation is measured. This study examines baseline sensitivity using published gate-all-around (GAA) transistor degradation data and a zero-curvature baseline reconstruction rule. In a primary dataset of 24 degradation curves, each curve yields one admissible zero-curvature root, shifting the median nominal exponent n = 0.550 to a reconstructed value 1/m = 0.346. Within an established phenomenological framework, this shift changes the inferred correlation parameter from χ = −0.820 to χ = −1.899. The apparent decrease in the local exponent with accumulated degradation appears in 19 of the 24 nominal curves but remains in only 5 curves after baseline reconstruction and is not robust under cross-fitted analysis. Across 26 eligible external nanowire curves (≥5 time points, out of 32 total curves), 19 curves yield a single admissible zero-curvature root. These roots remain numerically stable under single-point deletion (median relative change in m of 2.97% across 15 eligible curves) without systematically improving the secondary log–log fit. Numerical root stability therefore does not establish physical baseline validity. Consequently, the empirical exponent remains a useful kinetic descriptor, but it does not by itself distinguish between HCI- and BTI-like behavior or establish a transition between them.