In a half-bridge, the basic switching cell of solar inverters and DC-DC converters, two transistors take turns connecting an output node to the top or bottom supply rail. If both conduct at once, the supply is shorted ("shoot-through"), so controllers insert a short blanking gap, called dead-time, at each handover. The two handovers are not mirror images. The top switch's control signal must pass through a level shifter, a circuit that translates it to a voltage domain that floats with the output, and that path is measurably slower. The best gap is also set by different physics on each edge: one edge is a hard safety bound, while the other has an efficiency optimum. Every dead-time controller in the literature we reviewed treats the delays that set these gaps as fixed, or as depending only on load current. In reality they drift with transistor and driver temperature. Under rooftop solar, the converter's temperature lags its load by seconds to minutes while cloud shadows change the load in seconds. The same load current therefore occurs at very different temperatures during a cloud ramp.
We propose to treat the two edges differently. The efficiency edge can be tracked by measuring the switching node directly. The safety edge cannot be tuned that way, because you cannot observe a shoot-through constraint without approaching it. Its dead-time should instead be scheduled from a temperature-dependent delay model driven by an estimate of the converter's thermal state. The schedule itself is chosen offline as a Pareto set: the collection of best-available trade-offs among conversion loss, worst-case shoot-through margin, peak junction temperature, and output harmonic distortion. The claim is that, during realistic irradiance transients, this split controller either holds a safety margin that load-only schedules lose, or recovers efficiency that conservative fixed margins throw away. If it does neither, we will show that.
Core Claim and Why It Is Not Already Done
The closest work is by Sawan and colleagues. In A 9.2-ns to 1-s Digitally Controlled Multituned Deadtime Optimization for Efficient GaN HEMT Power Converters they built a digitally tuned dead-time chip. They showed that the low-to-high dead-time should simply be minimized subject to avoiding shoot-through, while the high-to-low dead-time has a load-dependent optimum set by zero-voltage switching. That optimum is the point where the output node's capacitance has resonantly discharged before the next switch turns on. Their chip is open-loop and indexed by load, and it gains 12% efficiency at 25 mA over a fixed 12 ns setting. Their companion design, An Active Dead-Time Control Circuit With Timing Elements for a 45-V Input 1-MHz Half-Bridge Converter, generates natively asymmetric delays of 35 ns and 62 ns because the high-side level-shifter path is slower. It cuts the high-side switch temperature from 141.4°C to 119.3°C, yet its own high-side delay swings between 50 ns and 90 ns across process corners. These two papers motivate the proposal. They establish that the edges differ and that delay drift is large, but neither lets temperature enter the timing decision, and neither is tested under a varying power source.
Two further results constrain the design. Circuit Techniques in GaN Technology for High-Temperature Environments found that GaN delay circuits and flip-flops fail at 160°C even though logic gates survive to 400°C. Timing blocks are therefore the temperature-fragile part, close to the 119–141°C range where these converters run. Separately, the transformer study Temperature Rise Calculation of Magnetic Core Considering the Temperature Effect of Magnetic Properties in an Electrical Steel Sheet shows that assuming constant, room-temperature material properties misestimated a hot spot by 29°C. It also shows that the fully coupled iterative solve is too slow to run inside a controller. This is why we compute schedules offline and keep the online part to a table lookup plus a lumped thermal estimate.
On the optimization side, AutoCircuit: Agentic Pareto Front Exploration for Analog Circuit Designs maps Pareto fronts for a low-power op-amp under static SPICE conditions. It does not address switching stages, temperature, or time-varying operation. The new element of this proposal is the combination of three things: the split between a sensed edge and a scheduled edge, temperature as a scheduling variable separate from load, and evaluation on real irradiance transients.
The safety condition we schedule against is the shoot-through margin
$$m(T, I) = t_{LH} - \big(\tau_{\text{off}}(T, I) - \tau_{\text{on}}(T)\big),$$
where $t_{LH}$ is the commanded low-to-high dead-time, $\tau_{\text{off}}(T,I)$ is the delay until the low-side switch has actually stopped conducting (a function of junction temperature $T$ and load current $I$), and $\tau_{\text{on}}(T)$ is the delay of the high-side turn-on path, including the level shifter. We require $m \ge m_{\min}$ at the worst process corner. Under the constant-property assumption $\tau_{\text{off}}$ and $\tau_{\text{on}}$ are evaluated at room temperature. We characterize their temperature dependence instead.
Research Program
The first stage builds an electro-thermal half-bridge model. It has three pieces. The first is temperature-dependent GaN switch behavior, covering on-resistance, the reverse-conduction voltage drop during dead-time, and the timing of output-node discharge. The second is temperature- and corner-dependent delay models for the dead-time generator and level shifter. These are calibrated to the 50–90 ns corner swing of the active controller and the 384 ps, 20% rise/fall asymmetry of A Novel Floating High-Voltage Level Shifter with Pre-Storage Technique. The third is a lumped thermal network, built in two versions: one with constant material properties and one with temperature-dependent conductivity and heat capacity. Coefficients come from bench characterization of an EPC2012-class buck stage, the platform of the multi-tuned work, and are not assumed. As a check, we will compare the model's switching-loss scaling with frequency against the measured sublinear factor of roughly 0.7 reported in Simulation-Based Power-Loss Optimization of General-Purpose High-Voltage SiC MOSFET Circuit Under High-Frequency Operation. That paper used SiC rather than GaN, so this is a sanity check only.
The second stage computes, on a grid of load current and junction temperature, the Pareto set of dead-time pairs. The objectives are per-cycle loss, worst-corner margin, and, for the inverter variant, output total harmonic distortion. Shoot-through margin behaves like a cliff, so we expect the loss–margin front to be non-convex. We will therefore use the hypervolume scalarization of Optimal Scalarizations for Sublinear Hypervolume Regret, which provably reaches concave parts of the front, and compare it with weighted sums. An operator picks one point per cell, and that choice becomes the schedule.
The third stage is the online controller. The high-to-low edge is tuned by closed-loop detection of when the output node finishes discharging. The low-to-high edge reads the temperature-indexed schedule using a junction-temperature estimate from the lumped thermal model, which is fed by measured current and optionally corrected by a slow on-die sensor. The fourth stage drives everything with irradiance traces. The BSRN ground-station series released with CrossViViT (Improving day-ahead Solar Irradiance Time Series Forecasting by Leveraging Spatio-Temporal Context, data via gitbooo/CrossViVit) are mapped to converter load through a fixed PV array model. We will separate the "Hard Days" at the Cabauw station, which have frequent cloud transients, from clear days. The inverter variant uses sinusoidal pulse-width modulation so that harmonic distortion can be compared against the dead-time-induced distortion studied in Analysis and Compensation Control of Dead-Time Effect on Space Vector PWM and An Improved SPWM Strategy for Effectively Reducing Total Harmonic Distortion.
Baselines
The naive reference is a conservative fixed asymmetric dead-time sized for the worst process and temperature corner, for example the 90 ns high-side corner of the active dead-time controller with a matching low-side setting. Nothing in this proposal can beat it on safety, so it anchors the efficiency cost of conservatism. Alongside it we run the conventional fixed 12 ns setting used as the reference in the multi-tuned work, and the fixed 35/62 ns asymmetric pair from An Active Dead-Time Control Circuit With Timing Elements for a 45-V Input 1-MHz Half-Bridge Converter.
The strongest existing method is the load-indexed multi-tuned schedule of A 9.2-ns to 1-s Digitally Controlled Multituned Deadtime Optimization for Efficient GaN HEMT Power Converters. It reports a 12% efficiency gain at 25 mA and 7% at 100 mA over the fixed 12 ns baseline on an EPC2012 GaN buck. We implement it on the same model and hardware. Its table is built by the same offline optimizer as ours, with the same compute budget, but with temperature held at a single calibration point. Our method beats it only if, over the Hard-Day traces, it holds worst-case margin at or above $m_{\min}$ wherever the load-only schedule violates it, with no loss in trace-averaged efficiency. The alternative way to win is to match the load-only schedule's margin while using less energy than the conservative fixed setting.
A second strong competitor is fully closed-loop adaptation on both edges, emulated by letting the low-to-high edge back off whenever the sensed overlap indicator rises. We give it the same sensing hardware model.
Two ablations isolate what is new. The first keeps our controller intact but replaces the temperature-dependent delay and thermal models with room-temperature constants. This is the constant-material-property ablation, and if it performs as well as the full controller, temperature scheduling adds nothing. The second replaces hypervolume scalarization with weighted sums at matched evaluation budgets and compares the resulting front hypervolume.
Why not just…?
Why not just use a large fixed dead-time and accept the loss? That choice is safe by construction, but the active-controller paper attributes a 40% reduction in dead-time-dependent losses, and a 22.1°C drop in switch temperature, to moving away from padded gaps. The fixed-conservative baseline will measure exactly how much trace-integrated energy and peak temperature this costs on Hard Days. If the cost turns out to be small at our operating point, we will report that the problem does not justify the controller.
Why not just extend the load-indexed lookup table, which already captures the efficiency optimum? This is the main threat. Under slow, steady load, temperature is nearly a function of load, so the load-only table implicitly sees temperature. Our hypothesis rests on thermal lag during cloud ramps breaking that relationship. The decisive test splits the evaluation into clear days and Hard Days. If the load-only schedule matches ours on Hard Days as well, temperature carries no extra information and the idea fails.
Why not just sense the switching node on both edges and adapt in closed loop? For the high-to-low edge we agree and adopt sensing. For the low-to-high edge, the quantity being protected is overlap conduction, and a sensing loop learns it only by approaching it. A process-and-temperature drift of 50–90 ns can also outpace a cautious adaptation rate. The closed-loop baseline tests this directly. If it holds margin through ramps without margin violations, the scheduled edge is unnecessary, and we would narrow the contribution to the Pareto characterization.
Why not just weight the objectives and minimize one number? Weighted sums cannot reach concave regions of a Pareto front (Optimal Scalarizations for Sublinear Hypervolume Regret). This matters only if the loss–margin front is actually non-convex here. The scalarization ablation measures whether it is. If the front is convex, we will use weighted sums and say so.
Decisive Checks
The primary metrics are four: trace-averaged conversion efficiency, minimum shoot-through margin over each trace, peak junction temperature, and, for the inverter, output harmonic distortion. Each is reported separately for clear days and Cabauw Hard Days, plus front hypervolume for the offline stage. The first decisive check is whether junction temperature at matched load differs enough during irradiance ramps to move the margin-optimal low-to-high dead-time by more than the controller's resolution. If it does not, the project stops there with a useful negative result. The second is the constant-property ablation versus the full model on Hard Days. The third is hardware confirmation of the margin result on a GaN buck stage, using recorded Hard-Day load profiles played back through an electronic load, with case temperature logged against model estimates. No reported result was found in the research file for dead-time controllers evaluated under solar irradiance profiles, so the load-only schedule run on our traces defines the reference figure.
References
- A 9.2-ns to 1-s Digitally Controlled Multituned Deadtime Optimization for Efficient GaN HEMT Power Converters<i/>
- An Active Dead-Time Control Circuit With Timing Elements for a 45-V Input 1-MHz Half-Bridge Converter
- Circuit Techniques in GaN Technology for High-Temperature Environments
- Temperature Rise Calculation of Magnetic Core Considering the Temperature Effect of Magnetic Properties in an Electrical Steel Sheet
- AutoCircuit: Agentic Pareto Front Exploration for Analog Circuit Designs
- A Novel Floating High-Voltage Level Shifter with Pre-Storage Technique
- Simulation-Based Power-Loss Optimization of General-Purpose High-Voltage SiC MOSFET Circuit Under High-Frequency Operation
- Optimal Scalarizations for Sublinear Hypervolume Regret
- Improving *day-ahead* Solar Irradiance Time Series Forecasting by Leveraging Spatio-Temporal Context
- Analysis and Compensation Control of Dead-Time Effect on Space Vector PWM
- An Improved SPWM Strategy for Effectively Reducing Total Harmonic Distortion