Beyond Exact Current Matching: Fabrication‐Tolerant Design of All‐Perovskite Tandem Solar Cells via Surrogate Modeling and Diode‐Based <i>J</i> – <i>V</i> Reconstruction

All‐perovskite tandem solar cells (TSCs) offer a promising route toward efficient photovoltaics, but their performance is affected by current mismatch and thickness variations. Here, we develop an explainable and robust optimization framework for a two‐terminal CsPbBr 3 /CsPbI 3 /MASnI 3 TSC. A coupled optical–electrical finite element method (FEM) generates an absorber‐thickness dataset to train a multioutput multilayer perceptron surrogate for predicting subcell J SC , V OC , fill factor (FF), and power conversion efficiency (PCE). The surrogate is integrated with nominal and Monte Carlo‐based robust optimization to identify thickness configurations that balance subcell currents while maintaining a high limiting current under ±10 nm variations. The nominal design achieves near‐perfect current matching of 0.009 mA/cm 2 in the prediction and 0.08 mA/cm 2 in the FEM validation, but its mismatch exceeds 2 mA/cm 2 under thickness variation. In contrast, the robust design accepts a finite mismatch of 0.516 mA/cm 2 in the prediction and 0.54 mA/cm 2 in the FEM validation, but occupies a flatter region of the design space. SHapley Additive exPlanations (SHAP) analysis interprets the learned relationship between thickness and current mismatch, showing that mismatch sensitivity arises from coupled spectral redistribution among subcells rather than independent absorber‐thickness contributions. Diode‐based reconstruction yields a PCE of 29.61% for the robust design versus 28.86% for the nominal case. These findings establish robust design as a reliable paradigm than strict current matching for high‐performance perovskite TSCs.