Physics-aware calibration neural network for automated TCAD-to-experiment modeling of carbon nanotube field-effect transistors

Abstract Low-dimensional semiconductor devices, including carbon nanotube field-effect transistors (CNT-FETs), require efficient calibration methods to reconcile numerical simulations with experimentally measured electrical characteristics. Conventional manual calibration is time-consuming, while purely data-driven surrogate models may provide limited physical interpretability and reduced reliability when applied to device geometries outside the training domain. In this work, we propose a physics-aware calibration neural network (PACNN) for supervised technology computer-aided design (TCAD)-to-experiment calibration of CNT-FET transfer and output characteristics. PACNN uses current–voltage curves generated by a two-dimensional intrinsic silicon TCAD model as a stable numerical reference. The architecture combines an anchor-preserving physical-affine module, which accounts for global variations in voltage alignment, current amplitude, slope, and extracted electrical parameters, with family-specific bounded residual modules that correct the remaining local discrepancies. Transfer characteristics are processed in the signed logarithmic current domain, whereas output characteristics are calibrated in the normalized linear current domain using separate amplitude and low-rank shape corrections. Model training incorporates curve-fitting, curve-level R 2 , slope consistency, device parameter, residual, and regularization objectives. Channel-length generalization was evaluated through out-of-distribution (OOD) extrapolation to $${L}_\mathrm{ch}$$ L ch = 4.5 μm and 5.0 μm. PACNN achieved macro-averaged curve-wise R 2 values of 0.947 4 and 0.876 0 for the transfer and output characteristics, respectively, with corresponding pooled pointwise R 2 values of 0.962 7 and 0.996 2. The predicted curves also preserved key electrical parameters, yielding a balanced physical consistency index of 0.918 3. These results indicate that PACNN provides an accurate and physically constrained framework for automated TCAD-to-experiment calibration of CNT-FET characteristics.