Sergio García-Sánchez

We investigate the frequency-dependent nonlinearities of an AlGaN/GaN high electron mobility transistor (HEMT) involved in terahertz (THz) detection by means of 2-D Monte Carlo (MC) simulations. The analysis shows that, below 1 THz, the responsivity roll-off is mainly governed by the passive microwave behavior of the device rather than by any limitation of the intrinsic detection mechanism. At higher frequencies, an inverse extraction method able to provide the intrinsic nonlinearity coefficients of the device reveals just a marginal broad enhancement around 1 THz, followed by a steep roll-off.

Self-heating significantly impacts the performance and reliability of GaN high electron mobility transistors, but capturing these effects with electrothermal Monte Carlo (MC) simulations is computationally intensive. This paper presents the application of a hybrid AI-thermal model, previously tested on another device, to the electrothermal analysis of GaN HEMTs. This first component consists of an artificial neural network (ANN) trained on isothermal MC data to predict drain current and lattice temperature. To extend the framework, a set of ANN-based microscopic models is introduced, composed of three dedicated networks that reconstruct spatially resolved quantities—electric field, carrier velocity, and sheet electron density. The system is coupled with compact thermal resistance models and iterated until convergence. The proposed approach achieves excellent agreement with electrothermal MC simulations while reducing computation time by approximately an order of magnitude. In addition to global performance metrics, it provides detailed internal profiles under electrothermally consistent conditions, making it a practical tool for fast device evaluation, in-depth analysis, and integration into compact modeling flows.

This study investigates the performance of planar Gunn diodes based on highly doped gallium nitride (GaN) using Monte Carlo simulations. The conversion efficiency is evaluated in geometrically V-shaped channels with an active region length of 500 nm, an input channel width of 200 nm, and output widths of 200, 220, and 250 nm. The diodes are subjected to various biasing conditions to assess dc-to-ac conversion efficiency under different ac biases, simulating operating conditions similar to those found in tuned circuits (comprising R, L, and C elements). The efficiency is analyzed for an ac voltage of 2 V superimposed on a 20-V dc bias, considering four distinct doping levels in the active region. These devices demonstrate conversion efficiencies of up to 0.36% at frequencies of 340 GHz for a channel doping level of N_D = 1.0 × 10¹⁸ cm^-3 and an output width of 250 nm. The increase in lattice temperature reduces the efficiency of the diodes, although the obtained values indicate that the devices would still remain operational. In addition, the frequency range where efficiency is positive (generation band) decreases as temperature increases.

Impact ionization originated by the buffer leakage current, together with high electric fields ( > 3 MV cm⁻¹) at the anode corner of the isolating trenches, has been identified as the failure mechanism of shaped planar GaN Gunn diodes when biased above 20 V, so that no evidence of Gunn oscillations in fabricated devices has been observed yet. In order to avoid the avalanche, we propose the addition of a Schottky substrate terminal, which, by means of Monte Carlo simulations, has been confirmed to be able to suppress such not-desired leakage current when applying a negative substrate bias. When the substrate bias is positive, impact ionization is also reduced due to the lower electric field at the hotspot, but a vertical cathode-substrate current degrades the device operation. In order to avoid such current, we propose the use a MIS configuration for the substrate contact, which is the optimal solution.

This article presents a hybrid artificial intelligence (AI)-thermal model for the determination of the current and lattice temperature of a device under a given bias voltage. The model is based on a neural network trained with isothermal Monte Carlo simulations and the coupling of any thermal model where the lattice temperature depends on the dissipated power. The proposed procedure has been validated on a gallium nitride (GaN)-based self-switching diode, although its application to other electronic devices, such as transistors, is also straightforward. The proposed method allows for a significant reduction in computational cost, in addition to enabling the investigation of various thermal models in an efficient manner. It is capable of reproducing the results that would be obtained through electrothermal Monte Carlo simulations, which are particularly computationally expensive.

An investigation into self-switching diodes based on highly doped GaN is conducted under direct current (dc) bias conditions. Different device geometries are explored under various lattice temperatures and polarization scenarios. Also, the impact of adopting an intervalley energy $\varepsilon _{{1}-{2}} =0.9$ eV for this material is examined and compared with results obtained with the traditionally accepted value of 2.2 eV. For a rectangular channel configuration, simulations predict oscillation frequencies in excess of 200 GHz, much above the expected transit-time value, due to the fact that the Gunn domains are formed near the anode side of the channel. Conversely, structures with a V-shaped geometry are able to start the formation of the Gunn domain inside the channel, thus generating oscillations at much lower frequencies (tens of gigahertz). The key result is that the lower $\varepsilon _{{1}-{2}}$ leads to smaller threshold voltage values (and also slightly smaller oscillation frequencies), particularly in diodes with short channels.

The responsivity of sub-THz zero-bias detectors based on GaN high-electron mobility transistors (HEMTs) is measured up to 110 GHz. A predictive model based on static coefficients extracted from the dc output curves together with S-parameter measurements is able to replicate the detection results both when the radio frequency power is fed into the drain or the gate terminals. The gate-drain coupling, counteracting the contribution of the direct modulation of the drain voltage to the RF responsivity, is at the origin of the frequency roll-off of the detection, and therefore, the frequency performance is improved when shrinking the gate size of the transistors. The extraction of the small-equivalent circuit allows understanding the influence of the parasitic elements and the scaling of the gate width of the transistors on the detection performance, as well as extrapolating the results of the measurements to predict the performances at higher frequencies. In addition, the analysis of the influence of the different equivalent-circuit elements allows linking the values of the responsivity and noise equivalent power (NEP) to the physics of the transistor operation and thus determining at which frequency the detection roll-off appears as a function of the gate length and width, and confirming that both must be reduced as much as possible to optimize the transistor high-frequency operation, the latter having a more significant influence.

In this work, by means of Monte Carlo (MC) simulations, the current responsivity of AlGaN/GaN HEMTs operating as zero-bias detectors is analyzed, reaching the THz frequency range. Two approaches are used for the calculation of the responsivity, trying to get physical information on the detection mechanisms from the comparison of their respective results. First, we determine the responsivity directly from the MC values of dc drain current shift originated by an input ac drain voltage excitation. As second approach, the responsivity is estimated from a closed-form expression involving the MC calculation of both the dc I–V curves of the transistor (to determine Taylor expansion coefficients) and the ac response in terms of the Y parameters (to be converted into S parameters). Both approaches coincide at low frequency, but the closed-form expression starts to deviate from the correct result at frequencies around 1 THz. Moreover, a modest plasma resonance is visible in the 2–5-THz range.

A generic matrix model for the responsivity of millimeter-wave (mmW) and sub-THz square law detectors is proposed. The model is valid for any number of ports and takes the biasing conditions into account, the power distribution, the phase shifts between ports, and the frequency dependence of the detector response. The proposed approach is experimentally validated up to 67 GHz on an AlGaN/GaN high-electron-mobility transistor (HEMT). It could, however, be used for other types of detectors, including diodes and CMOS detectors. In addition, a reliable mathematical method is presented, allowing the direct determination of the maximum attainable responsivity and the associated optimum excitation. This method avoids a computing time-consuming evaluation of the above quantities by constrained optimization and thus provides an efficient way to evaluate the performances of a square law detector only based on its dc and $S$ -parameters characterization.

The behavior of GaN-based high-electron-mobility transistors (HEMTs) as microwave zero-bias detectors is very dependent on the configuration of the bias (current or voltage), the operation temperature, and whether the radio-frequency power is fed in the drain or the gate terminal. When the signal is injected into the drain, the negative current responsivity shows a bell-shape dependence on $V_{\rm GS}$ centered slightly above the threshold voltage for all the studied frequencies (1–43 GHz) and temperature ranges (8–400 K). In the case of the voltage responsivity, depending on the temperature range (associated with the presence or absence of drain leakage current), an increase or decrease of the responsivity is observed in subthreshold conditions. For the gate-injection (GI) configuration, as expected by the capacitive gate–drain coupling, the voltage responsivity at low frequency is null, but only for $V_{\rm GS}$ above threshold. Surprisingly, in subthreshold conditions, it is very high and positive, contrary to the negative values intuitively expected for this configuration. The origin of this unexpected behavior, taking place in both gate and drain injection (DI) configurations, is that the drain terminal is self-biased at the zero-current (ZC) operating point ( $V_{\rm DS}$ being largely negative). An analytical model based on static coefficients obtained from dc measurements can explain the mechanisms behind the observed dependencies of the experiments.