The identification of the microscopic defects responsible for charge trapping will be a major key to find recipes to improve the reliability of future GaN technology. The results derived throughout this work should thus be seen as a basis for further studies in the field of the reliability of GaN technology. The following paragraphs list some ideas for future research based on the findings above.
Although several types of bulk defects affecting the reliability of GaN have been found by measurements and are widely accepted throughout the community, one of the most crucial defect(s) necessary for HEMT devices, the surface donors, remain largely unknown. To identify the structure and the origin of these defects, future studies should combine the extraction of defect bands from large area devices with single-defect experiments and first-principle simulations.
From a technological point of view, this requires the fabrication of large-area devices together with nano-scale devices in the same process to ensure compatibility. Furthermore, the device layout should be kept as simple as possible (i.e., planar devices) to simplify the calibration of the device in TCAD simulations. For large-area devices, the feedback of charges captured in defects plays a crucial role. This should be taken into account already when designing eMSM measurements, but also in device simulations, as the observed degradation will be a function of the stress history of the device.
As the devices usually suffer from large instabilities already at nominal operating conditions, measurements at cryostatic temperatures allow to reduce thermal noise and trigger less response of defects. For single-defect measurements, the barrier layer should be designed as thin as possible in order to raise the average step-heights of defects at the interface to a level which is detectable by the equipment. The usage of constant-current setups can also help to detect defects with smaller step-heights as the full measurement resolution is available across the whole current range. Alternatively, MIS-HEMTs with small oxide thicknesses could be used to monitor the gate current which in turn allows defect modeling based on the observed leakage currents.
The impact of the barrier on the observed time-constants of the surface donors is another topic, which is widely neglected in reliability investigations of GaN HEMTs up to date. Separating the influence of the barrier from native defect properties, however, could be one of the keys to identify the origin and structure of the surface donors.
The HMM library was so far only tested on real measurement data from GaN technology. Although it should be independent of the technology the data was recorded on, a broader set of tests should help to reveal eventual bugs and further improve the general robustness of the algorithms.
The first and most valuable improvement would be the implementation of finite HMMs, which would also enable the processing of TDDS data. Currently, the HMM does not consider an explicit state which ends a sequence. In the case of TDDS data, such an explicit end-state would naturally be given when all defects emitted their charge. The average occupancies of the defects after stress can be directly obtained from the starting probabilities of the HMM after training.
Another major improvement would be the implementation of a factorial HMM, which limits the size of the system from being the factor of the states to the addition of the states of the underlying defects. A state space of a system of three three-state defects thus would be reduced from 
to 
. This would allow to use a larger number of defects at the expense of complicating the addition of thermal states.
One obvious improvement is related to the training speed of the library. In the current state, about half of the running time is dedicated the baseline estimation which cannot be processed in parallel yet. A clever parallelization algorithm of this part of the library or faster baseline estimation algorithms would dramatically improve run-times. Moreover, the baseline estimation algorithms themselves need to be optimized further in terms of both, their speed and robustness against different measurements.
Up to date, the Baum-Welch algorithm to find the sequence of states is independent of the baseline estimation algorithm. Combining these two parts into a single expectation maximization algorithm for the combined system including the baseline could help to improve both, the convergence and the quality of the training results. Further, this algorithm could also be designed to add other constraints to the Baum-Welch algorithm which prohibit the necessity of splitting up and re-assembling the system of defects at every iteration. A more formal inclusion of these constraints could also help to judge on their impact on the global training results, which is largely unknown up to this point.
