Inverse design of 2D metal halide photodetectors based on machine learning

Two-dimensional (2D) metal halide photodetectors offer great potential for high-performance optoelectronic applications. However, the design of these devices remains challenging because responsivity and detectivity are controlled by coupled material, structural, contact, and operating descriptors. Here, based on a literature-derived dataset comprising 224 device-performance records, we propose a physics-constrained inverse-design framework that integrates machine-learning prediction, SHAP (SHapley Additive exPlanations), domain-knowledge-guided causal-effect estimation, principal component analysis (PCA)-based applicability-domain assessment, and finite-element simulation. XGBoost and Random Forest (RF) are identified as the optimal surrogate models for responsivity and detectivity, respectively. SHAP analysis highlights Bandgap as the dominant predictive descriptor for both targets. DAG-guided g-computation estimates target-dependent total intervention effects under exposure-specific back-door adjustment and identifies physically plausible pathways involving electronic-structure, contact-related, and spacer-related descriptors. The inverse-design process yields ten chemically admissible descriptor-level configurations toward the high-performance region of the original descriptor space, whose device-level performance is subsequently evaluated computationally through finite-element simulations. PCA places the candidates near the sparse high-performance boundary. This study provides an interpretable and physics-constrained machine learning paradigm for the data-driven design of high-performance 2D metal halide photodetectors.