Machine learning-driven optimization of the tunability—Efficiency trade-off in dual-modulated MoS2/WSe2 van der Waals optoelectronics

In van der Waals heterostructures, the quantum-confined Stark effect enables electrical control of optical transitions, yet it carries an intrinsic penalty: every field increment that redshifts the emission also pulls the electron and hole wavefunctions apart, eroding the oscillator strength on which device efficiency depends. We report a dual-modulation strategy for MoS2/WSe2 bilayers that pairs biaxial tensile strain (0%–2.0%) with a reduced perpendicular electric field (0–200 kV/cm). The strain pre-conditions the permanent interlayer dipole moment, so a smaller field achieves a given spectral shift while wavefunction separation stays limited. We assembled a 5000-point density functional theory with Grimme's D3 dispersion correction and Becke–Johnson damping [DFT-D3(BJ)]/Perdew–Burke–Ernzerhof exchange-correlation functional (PBE) and technology computer-aided design dataset by Latin hypercube sampling and trained a deep neural network surrogate (R2 > 0.98); it outperformed Gaussian process, random forest, and polynomial response-surface baselines, most clearly in the super-additive nonlinear interaction regime. NSGA-II optimization mapped the full Pareto frontier. At ɛ = 1.5% and F = 85 kV/cm, the balanced operating point delivers a 340 nm redshift at a normalized oscillator strength of 0.48, a 218% gain over the field-only reference at the same shift. First-order perturbation analysis attributes this to a strain-induced reduction of the dipole moment from 0.42 to 0.31 eÅ, which lowers the required field by 43% and reduces wavefunction separation in proportion. HSE06 benchmarking of five Pareto configurations confirms that relative PBE trends hold to within 6%. Under ±5% fabrication tolerances, Monte Carlo simulation gives an efficiency variation of 3.3% for this configuration, against 41% for the field-only device, a 12.5-fold gain in process robustness.