At one point during this project, we had a newly developed photodetector, a box of silicon wafers, and a question that was not really part of our original experimental plan: how many silicon wafers could we actually see through?
Silicon looks completely opaque to our eyes, but at short-wave infrared (SWIR) wavelengths around 1300 nm, it becomes sufficiently transparent for light to pass through. We had already been using this property in our imaging experiments, so we decided to push it a little further. We placed a resolution target behind the silicon and began adding wafers, each approximately 0.5 mm thick. With each additional wafer, the finer features became less distinct, while the larger patterns remained visible. We kept going through five wafers, ten, fifteen and twenty, until eventually we reached 24 wafers, corresponding to a total silicon thickness of approximately 12 mm. By then, the image was noticeably blurred and washed out, but some of the larger features of the target could still be recognized.
[How far could we push it? SWIR images of a USAF resolution target acquired through an increasing number of silicon wafers. With sufficiently strong illumination, larger features remained recognizable even through a stack of 24 wafers.]
The 24-wafer experiment was never intended to be a performance benchmark. We had increased the illumination simply because we were curious to see how far the setup could be pushed. In fact, the detector had been developed with almost the opposite situation in mind: obtaining useful information when the available optical signal is weak. For the experiment that eventually became part of our Nature Communications paper, we therefore returned to more demanding low-light conditions and used a stack of six silicon wafers. At that point, the question was no longer how many wafers we could see through, but whether the visible and SWIR capabilities of a single detector could help solve a practical imaging problem.
From seeing through silicon to aligning it
That problem comes from the increasing use of vertically integrated semiconductor structures. Instead of placing every component side by side, modern devices can stack multiple semiconductor dies and connect them vertically. High-bandwidth memory is one familiar example. In these structures, through-silicon vias, or TSVs, provide vertical electrical connections through the silicon layers. For those connections to work reliably, the layers have to be accurately aligned before they are permanently bonded.
This creates a surprisingly straightforward imaging question: are the alignment features on opposite sides of a silicon wafer in the right positions relative to one another? Features on the front side can be observed using reflected visible light, whereas features behind the silicon can be accessed using SWIR light transmitted through the substrate. The difficulty is that visible and SWIR images are often acquired using separate sensors or different optical paths. Once two imaging systems are involved, their own relative alignment becomes part of the measurement, introducing an additional calibration step and another possible source of error.
Our device offered a simple way to approach this problem. The photodetector combines an organic photoactive layer and a PbS quantum-dot layer in a single monolithic structure, and its spectral response can be switched simply by changing the applied bias. At +1 V, it operates in the visible mode, while reversing the bias to −1 V switches the detector to the SWIR mode. The detector itself does not have to be replaced or moved. This means that visible and through-silicon SWIR information can be acquired from the same sensing position, avoiding the need to register two independent detectors.
[One detector, two views. By electrically switching the spectral response of our monolithic organic/PbS photodetector, we acquired a visible reflection image from one side of the sample and a SWIR transmission image through the silicon stack. Corresponding alignment features in the two images were then used to estimate rotational and translational offsets.]
To test the idea, we built the proof-of-concept setup shown above. Dot-pattern masks were positioned on opposite sides of a stack of six silicon wafers to emulate alignment features. The detector acquired one pattern in the visible mode and the other using 1300-nm SWIR light transmitted through the silicon stack. Importantly, the detector remained in the same position between the two measurements; we changed only the applied bias and, consequently, the spectral response of the device.
We then deliberately introduced rotational and translational misalignment so that we could test whether the displacement could be identified and corrected. Rather than using every corresponding marker to calculate the transformation, we divided the points into two groups. One group was used to estimate the geometric correction, while the remaining points were withheld from the fitting process and used afterward to evaluate whether the estimated transformation also worked on independent markers. Following a physical correction corresponding to a −9.43° rotation together with sub-millimeter translation, the mean residual offsets were reduced to −5.63 μm along the x direction and −4.79 μm along y.
[Proof-of-concept alignment through a six-wafer silicon stack. Visible and SWIR images acquired using the same bias-switchable detector are used to identify corresponding alignment features, estimate the rotational and translational displacement, and guide physical correction.]
For us, the interesting part of this experiment was not only the final alignment error. The more important point was how the two pieces of information were obtained. Instead of acquiring the visible image with one detector and the SWIR image with another and then calibrating the two systems against each other, both measurements came from the same stationary device. The spectral window was selected electrically.
There is, of course, a considerable distance between this proof of concept and a practical semiconductor inspection system. Our alignment mask was 35 mm × 35 mm, and the individual markers were still far larger than real TSV structures. The images were also reconstructed by mechanically scanning a single-pixel detector across the field of view, which is useful for demonstrating the sensing principle but not suitable for high-throughput inspection.
The next challenge is to translate this proof of concept toward a practical imaging platform: moving from a mechanically scanned single detector to an imaging array, reducing the scale from millimeter-sized test patterns to microscopic alignment features, and ultimately evaluating the approach on realistic semiconductor structures. Although substantial development is still required, our results suggest that electrically switchable multispectral detection could provide a simple way to acquire visible and through-silicon information from the same sensing platform.