UCSC-SOE-09-24: On the Origin of On-Orbit Upset Uncertainties
Robin D. Morris, Charles C. Foster, Athanasios Kottas, and Marian Farah
For parts that are to be used on-orbit, the final result of a set of
heavy-ion tests is a prediction of the upset rate for a particular orbit. This prediction is made by combining the cross-section vs. LET for the part with a model of the on-orbit radiation environment. Typically a Weibull curve is used to model the cross-section vs. LET. Due to limited testing, there will be uncertainty in determining the Weibull parameters from the test data, and hence uncertainty in the cross-section vs. LET response. This results in uncertainty in the predicted on-orbit upset rate. We present statistical methodology that can capture these uncertainties. We extend the framework to allow for the inclusion of other sources of uncertainty – in the fluence recorded during the test; the actual LET in the sensitive volume (due to overlayers or testing through the back of the die); the beam uniformity; the device thickness; and other test/device parameters. We analyze in detail the effect of a 5% uncertainty in the fluence, and confirm the received wisdom that fluence uncertainty can be neglected, though the reasons are not as simple as may be expected. The use of a Weibull to model the cross-section vs. LET response is purely conventional; it has no physical justification. Other models (e.g., log-normal) can be used, and the resulting model uncertainty has an impact on the uncertainty of the on-orbit upset rate predictions. We fit a log-normal curve to the test data, and show the impact of model choice on the on-orbit upset rate distribution. We also use a non-parametric estimate of the cross-section vs. LET, which makes much weaker assumptions about the form of the cross-section vs. LET. The parameters of the Weibull curve are ill-determined if the test data does not reach the plateau. We demonstrate how our statistical framework allows this case to be treated, and produce estimates of the parts’ on-orbit upset rate distributions. Using the uncertainty in the predicted on-orbit upset rates as a quality measure can give valuable information regarding when sufficient testing has been performed.
For parts that are to be used on-orbit, the final result of a set of
heavy-ion tests is a prediction of the upset rate for a particular orbit. This prediction is made by combining the cross-section vs. LET for the part with a model of the on-orbit radiation environment. Typically a Weibull curve is used to model the cross-section vs. LET. Due to limited testing, there will be uncertainty in determining the Weibull parameters from the test data, and hence uncertainty in the cross-section vs. LET response. This results in uncertainty in the predicted on-orbit upset rate. We present statistical methodology that can capture these uncertainties. We extend the framework to allow for the inclusion of other sources of uncertainty – in the fluence recorded during the test; the actual LET in the sensitive volume (due to overlayers or testing through the back of the die); the beam uniformity; the device thickness; and other test/device parameters. We analyze in detail the effect of a 5% uncertainty in the fluence, and confirm the received wisdom that fluence uncertainty can be neglected, though the reasons are not as simple as may be expected. The use of a Weibull to model the cross-section vs. LET response is purely conventional; it has no physical justification. Other models (e.g., log-normal) can be used, and the resulting model uncertainty has an impact on the uncertainty of the on-orbit upset rate predictions. We fit a log-normal curve to the test data, and show the impact of model choice on the on-orbit upset rate distribution. We also use a non-parametric estimate of the cross-section vs. LET, which makes much weaker assumptions about the form of the cross-section vs. LET. The parameters of the Weibull curve are ill-determined if the test data does not reach the plateau. We demonstrate how our statistical framework allows this case to be treated, and produce estimates of the parts’ on-orbit upset rate distributions. Using the uncertainty in the predicted on-orbit upset rates as a quality measure can give valuable information regarding when sufficient testing has been performed.