Accelerated Burn-In Process for High Power Quantum Cascade Lasers to Reduce Total Cost of Ownership

Despite their widespread adoption, one of the main hurdles holding QCLs back from large volume manufacturing is the large cost of ownership. While QCLs, like most semiconductor devices based on III-V compounds, can leverage the economies of scale typical of semiconductor manufacturing, and therefore lower production cost at wafer and chip level, it is at the testing and packaging stages of QCL production that most of the cost can build up and not scale as easily in large volumes. One of the most time-consuming steps during the post-fabrication testing and packaging process is device burn-in. Burn-in generally consists in running the devices in controlled conditions deemed representative of the actual operating specs in the field for a number of hours. This is done in order to screen out possible early degradation issues that may lead to costly device replacement after system installation is complete. Notwithstanding the importance of this issue, up till now there are no recognized industry standards for device burn-in qualification procedures and there are no proven models for laser degradation to support such procedures. Optimizing early screening of defective devices by improved burn-in procedures is likely going to reduce post-fabrication costs (testing, packaging, installation, etc.) which in large volume manufacturing typically represent the largest portion of production costs. A long burn-in procedure (i.e. 100hrs or more) can help reduce the rate of rejection of QCLs from ~20% straight out of the wafer to less than 2% when delivered to the customer, so a 10-fold reduction in returned product costs both for the manufacturer and for the final user. This result though, comes at the cost of a very time-consuming procedure that cannot easily be scaled up to large volumes without increasing the cost of capital equipment and production logistics. Reducing the burn-in time to a few hours, or maybe even minutes, at least for the devices most likely to fail, will unlock a great cost reduction for the manufacturer that can be passed on to the final user. Instead of a top-down approach where the device behavior is explained by a priori assumptions about the degradation mechanisms and origin of defects, in this project we took a bottom-up approach where we trained machine learning (ML) algorithms to classify device characteristics based on the BI degradation results observed and built a predictive tool for future device behavior that can identify faulty devices from the very first measurements. This can drastically reduce the need for a long, time consuming burn-in to screen out defective devices. A significant reduction of testing cost and replacement cost by correct identification of defective devices at the early stages of their characterization will lead to significant savings for the customer in terms of both acquisition and maintenance costs. In Phase I we have shown that our predictive tool allows us to do just that.