Physics-informed sparse modeling for GaN devices

SparseMind and MRDI are studying sparse surrogate models for GaN device simulation. The programme covers GaN HEMTs and PIN diodes and uses physics-informed neural networks (PINNs) to incorporate physical constraints and reference data into training.

Method

The teams will align device geometry, operating conditions and TCAD reference data, establish training and evaluation baselines, then assess sparse models for accuracy, robustness and computational requirements. TCAD supplies reference data and validation results.

Acceptance targets

Targets include threshold-voltage error below 3%, dynamic on-resistance error below 5% and peak junction-temperature mean absolute error below 3°C. Subject to the voltage and resistance constraints, the model targets at least 60% sparsity, with at least 80% as a stretch target.

The team will compare results against TCAD under defined operating conditions and measure computational requirements for the target workflow.

Next steps

The research aims for 1.5–2× speedup and millisecond-scale surrogate predictions. Accuracy, sparsity, speed and latency remain to be validated on the target hardware, data and workload. The team will also investigate combining quantization with sparsity.