Surrogate-Guided Constrained Optimization of Ultra-Scaled MoS 2 FET Designs under Strict Manufacturability Constraints

Designing ultra-scaled MoS 2 field-effect transistors (FETs) requires navigating a high-dimensional design space under competing performance objectives and stringent manufacturability constraints. While random search and large language model (LLM)–based heuristics can propose candidate device parameters, their sample efficiency and feasibility under strict regimes remain limited. This paper presents a fully reproducible, end-to-end pipeline that integrates (i) a physics-inspired compact oracle for key device metrics, including on-current (I on), leakage ratio (I off /I on), subthreshold swing (SS), and drain-induced barrier lowering (DIBL); (ii) leak-free surrogate modeling with group-based data splitting; and (iii) surrogate-guided constrained optimization that iteratively proposes candidates and validates them via oracle evaluation. We formulate device design as a constrained, utility-scalarized multi-objective optimization problem and evaluate competing methods under both relaxed and strict manufacturability tiers. Across multiple random seeds, surrogate models achieve high predictive fidelity, with group-split R 2 values up to 0.99 for multiple targets. Surrogate-guided search consistently outperforms oracle random baselines, a Bayesian optimization baseline, and an optional LLM self-refinement heuristic. Under strict constraints, surrogate guidance substantially improves feasibility rates and identifies higher-utility designs with significantly fewer oracle evaluations. These results demonstrate the effectiveness of surrogate-guided strategies for sample-efficient constrained nanoelectronic design in controlled testbeds.