A Pragmatic Model to Predict Future Device Aging
A Pragmatic Model to Predict Future Device Aging
复制标题
DOI：
10.1109/access.2023.3329077
复制
发表时间：
2023
期刊：
影响因子：
3.9
通讯作者：
James Brown;K. H. Tok;R. Gao;Zhigang Ji;Weidong Zhang;J. Marsland;T. Chiarella;J. Franco;B. Kaczer;D. Linten;Jian Fu Zhang
中科院分区：
文献类型：
--
作者：
James Brown;K. H. Tok;R. Gao;Zhigang Ji;Weidong Zhang;J. Marsland;T. Chiarella;J. Franco;B. Kaczer;D. Linten;Jian Fu Zhang
为了预测正在使用偏差的长期设备老化，必须将从电压加速测试中提取的模型推断到将来。传统模型使用功率定律，将测试数据线性拟合在日志图中，然后推断衰老动力学。挑战在于，测得的数据并不总是遵循日志图图上的一条直线，将这种预测的准确性提出了质疑。尽管在这种情况下有一些模型可以很好地拟合测试数据，但它们对未来衰老的预测能力通常未验证。这项工作的关键进步是开发一种方法，用于提取模型，该模型可以在衰老动力学不遵循简单的功率定律时，可以清楚地预测广泛（VG，VD）偏置空间的未来衰老。这是通过将老化分为四种类型的陷阱并通过单独的直线建模的四种陷阱来实现的。在3个不同的CMOS过程中验证了该方法的适用性，在该过程中，它可以预测未来至少3个数量级的老化。映射了（VG，VD）空间的每种类型的陷阱的贡献。还表明，使用测试数据的良好拟合并不能保证良好的预测，因此不应将良好的拟合用作验证模型的唯一标准。
To predict long term device aging under use bias, models extracted from voltage accelerated tests must be extrapolated into the future. The traditional model uses a power law, to linearly fit the test data on a log-log plot, and then extrapolates aging kinetics. The challenge is that the measured data do not always follow a straight line on the log-log plot, calling the accuracy of such prediction into question. Although there are models that can fit test data well in this case, their prediction capability for future aging is typically not verified. The key advance of this work is the development of a methodology for extracting models that can verifiably predict future aging over a wide (Vg, Vd) bias space, when aging kinetics do not follow a simple power law. This is achieved by experimentally separating aging into four types of traps and modelling each of them by a straight line individually. The applicability of this methodology is verified on 3 different CMOS processes where it can predict aging at least 3 orders of magnitude into the future. The contributions of each type of traps across the (Vg, Vd) space are mapped. It is also shown that good fitting with test data does not warrant good prediction, so that good fitting should not be used as the only criterion for validating a model.