Fully Automated Quantum Cascade Laser Design Aided by Machine Learning with up to 100X Design Cycle Time Reduction
Navy STTR 2020.A - Topic N20A-T003
NAVAIR - Ms. Donna Attick donna.attick@navy.mil
Opens: January 14, 2020 - Closes: February 12, 2020 (8:00 PM ET)


TITLE: Fully Automated Quantum Cascade Laser Design Aided by Machine Learning with up to 100X Design Cycle Time Reduction


TECHNOLOGY AREA(S): Air Platform, Electronics


OBJECTIVE: Develop a fully automated Quantum Cascade Laser (QCL) design process by using neural networks and machine learning (ML) algorithms that will result in up to 100 times reduction in design cycle time compared to the conventional “manual” QCL design process [Ref 1].

DESCRIPTION: The active region of a QCL consists of many (typically 20-50) repeated stages of superlattice (SL) material. The highest-performance QCLs operating in the mid-infrared spectral region (approximately 4.8 micron) utilize an indium phosphide (InP) substrate and have active regions wherein each stage consists of 10s of ultrathin layers of indium gallium arsenide (InGaAs) quantum wells and aluminum indium arsenide (AlInAs) barriers. The device performance metrics (such as emission wavelength, threshold-current density, slope efficiency, and their temperature dependence) are closely tied to the quantum-confined state energies and their electronic wave-function spatial distributions within the active region, which in turn are determined by the specific layered structure (i.e., layer thicknesses and compositions). The complexity of the layered structure generally requires a time-consuming iterative process between experiment and design optimization to achieve the highest device performance, which adds substantial cost to QCL manufacturing. Automated optimization algorithms [Ref 1] applied to QCL design could both greatly reduce the time (and cost) required to develop new QCL devices with specified performance characteristics and potentially lead to new insights into QCL design.

The current QCL design process generally involves a human in the loop - even for a single iteration. The function performed by the human is to identify specific features in the design and determine whether a certain performance metric can be achieved. Emerging data-driven automated optimization algorithms could potentially address the difficulties facing QCL design.

As the QCL’s structural complexity grows, the design processes become more challenging. With conventional design approaches, based on computational optimization, one typically starts with a prior design and computes the performance, compared to the target response. The gradient of structural change in layers and compositions is calculated and applied to the design. This process, performed iteratively, often takes hundreds of iterations before a design is found that meets the design criteria. As an alternative, the data-driven approach is rapidly emerging where deep neural networks are used for inverse device design. A large data set of existing designs and corresponding performances can be used to train artificial neural networks so that the networks can develop intuitive connections between QCL designs and their performances. After training, the neural network can accomplish a design goal in hours instead of weeks as compared to the conventional approach. Such an approach has been used previously in photonic structures [Ref 2], where neural networks successfully model the wave dynamics in the Maxwell’s equations.

Demonstrate and deliver a single-mode QCL prototype that is designed using the algorithm(s). The specifications of the QCL for this design algorithm demonstration are:  > 15 W continuous wave output power at room temperature; M2 no more than 1.5 in both the fast and slow axes; laser emission in the spectral range between 4.6 to 5.0 micrometers; and no unexpected and undesirable beam steering effect as the QCL drive current is increased. Furthermore, the contractor is required to deliver the fully automated QCL design algorithms with complete and detailed user manual and documentations.

PHASE I: Develop a methodology for implementing the training plan for neural network-based QCL design optimization without human intervention. Establish performance metrics, including but not limited to, output power, beam quality, wall-plug efficiency, and thermal impedance, etc. The design verification plan for the algorithms will be implemented in Phase II. The Phase I effort will include prototype plans to be developed in Phase II.

PHASE II: Demonstrate fully automated QCL design algorithms using ML methodology. Perform experimental verification of the generated designs by demonstrating that the QCL performance metrics are met with less than +/- 2% variations from the target performance specifications. Demonstrate and deliver a single-mode QCL prototype that meets the design specifications. Deliver the fully automated QCL design algorithms with complete and detailed user manual and documentations. Benchmark the design cycle time using the algorithm aided by ML against the conventional method without using ML, and verify the cycle time reduction.

PHASE III DUAL USE APPLICATIONS: Test and finalize the technology based on the design and simulation results developed during Phase II. Transition the design algorithm for DoD applications in the areas of Directed Infrared countermeasures, advanced chemicals sensors, and Laser Detection and Ranging. Commercialize the design algorithm based on ML for law enforcement, marine navigation, commercial aviation enhanced vision, medical applications, and industrial manufacturing processing.


1. Bismuto, A., Terazzi, R., Hinkov, B., Beck, M. and Faist, J.  “Fully Automatized Quantum Cascade Laser Design By Genetic Optimization.” Applied Physics Letters, 2012. https://aip.scitation.org/doi/citedby/10.1063/1.4734389

2. Liu, D., Tan, Y. and Yu, Z. “Training Deep Neural Networks for the Inverse Design of Nanophotonic Structures.” Department of Electrical and Computer Engineering, University of Wisconsin: Madison, WI.  https://arxiv.org/ftp/arxiv/papers/1710/1710.04724.pdf

KEYWORDS: Mid-Infrared; Quantum Cascade Lasers; Infrared Countermeasures; Cycle Time Reduction; Machine Learning; Design Algorithm; ML; QCL