AI-Driven Surrogate Modeling and Inverse Design of Type-II Superlattice Infrared Detectors

Abstract

The conventional design of Type-II superlattice (T2SL) infrared detectors relies on iterative physics-based simulations, which can make broad exploration computationally costly. Here, we develop an AI-driven surrogate workflow for InAs/GaSb T2SL detectors. Two separately trained Transformer regressors with the same layer-token architecture predict dark-current-density curves and quantum-efficiency (QE) spectra from layer thicknesses, doping concentrations, and operating temperature. On fixed held-out test sets, the dark-current model achieved a root mean square error (RMSE) of 0.7332 A/cm2, mean absolute percentage error (MAPE) of 2.7913%, and coefficient of determination (R2) of 0.998103, while the QE model achieved an RMSE of 0.005104, MAPE of 1.6417%, and R2 of 0.999231. Controlled comparisons on identical data splits showed that the Transformer obtained the lowest RMSE and MAPE and the highest R2 for both targets relative to multilayer-perceptron and Random-Forest baselines. Using the two surrogates, we evaluated one million uniformly sampled candidates at 77 K and identified 83 structures meeting the target requirements: dark-current density below 5 × 10-6 A/cm2 at -1 V and mean QE above 56% over wavelengths from 3.03 to 3.42 μm. Six selected candidates were subsequently compared with Crosslight simulations. By combining surrogate prediction, high-throughput target-driven screening, and Crosslight simulation verification, this framework shifts T2SL detector design from repeated trial-and-error simulation toward intelligent prediction and verification, offering a practical route to accelerate the exploration of high-performance infrared detector structures.