It’s ISSCC week, and I got to see some interesting things, some of which go beyond interesting and possibly to significant. My focus was on MEMS, sensors, and emerging technologies. I’ll be following up on a number of bits and bobs here over the next several days.
In the MEMS/sensor realm, as was the case last year (where we did an entire series of sensor articles), the focus at ISSCC was on the conditioning circuits for sensors rather than the sensors themselves. The problems they solve are ever better resolution or accuracy with as little calibration as possible.
Just to start things off, the first of the papers dealt with a source of drift (or extremely low-frequency noise, since it’s reversible) in accelerometers. Most such compensations deal with non-idealities in the sensor or in the conditioning circuits themselves. But this picked on something that you might think would be a complete non-issue: the change in the capacitance contributed by the bond wires connecting the accelerometer sensor to its companion ASIC (given that they’re mostly not collocated on the same silicon) due to mechanical stress and humidity.
The reason you might not consider this to be significant is the simple fact that a packaged part usually has bond wires that are encapsulated in plastic, limiting their movement. In fact someone raised that as a question (since the experiments leading to the paper were done with exposed bond wires that could be manually deflected), and, in fairness, the paper doesn’t address how much deflection might actually be realistic. They do say, however, that a 1-µm deflection can contribute a 14-mG offset (“G” here being the gravity/acceleration G, the kind you “pull,” shown as “mg” in the paper, which looks to me too much like milligrams). A 13-µm deflection takes the offset all the way to the maximum tolerable error of 100 mG.
The other issue is absorption of moisture by the package, which changes the permittivity of the material, and hence the capacitance.
Most other sources of systematic error due to design and production can be calibrated out in the factory. These, however, cannot be. So the paper presents a means of doing this.
The main challenge is isolating only the capacitance of the bond wires and the resulting error. The approach they took was to modulate it up into a region of the spectrum that was within the low-noise portion but outside the operating signal band. They did this by applying a time-varying feedback force, measurements from which allowed them to provide a correction to the error.
A classic accelerator would have two electrostatic elements, one for forcing and one for sensing. But to save space, they (or, at least, this one) had only one. So they had to time-multiplex the correction with the normal operation of the sensor to provide continuous correction.
You can find details of the approach in Paper 11.1.