Machine‐Learning Surrogate Modeling of RbGeI 3 Perovskite Solar Cells: Global Sensitivity Analysis and the Limits of Surrogate‐Guided Optimization

Here, a machine‐learning surrogate is developed for an FTO/TiO 2 /RbGeI 3 /Spiro‐OMeTAD/Au configuration from a Sobol‐sampled dataset of 4396 SCAPS‐1D simulations across seven absorber parameters. Comparison of model architectures demonstrates that power conversion efficiency (PCE) prediction accuracy, measured by R 2 , improves from 0.764 for ridge‐regularized linear regression to 0.879, 0.927, and 0.968 for second‐degree, third‐degree, and gradient boosting formulations, respectively. Sobol sensitivity analysis attributes 60.5% of PCE variance to absorber defect density as a first‐order effect and 17.8% to higher‐order parameter interactions. Absorber thickness and acceptor doping show near‐zero first‐order but total‐order indices of 2.5% and 6.9%, their effects reverse sign across the domain, so a single one‐at‐a‐time sweep cannot fully characterize them. All candidate configurations, from joint and one‐at‐a‐time optimization, were generated on the surrogate and then simulated in SCAPS‐1D, which showed a 0.9‐percentage‐point over‐prediction of the surrogate's joint optimum. After identical refinement within SCAPS‐1D, the best one‐at‐a‐time candidate reached 27.39% within the search restrictions (thickness ≤ 0.8 µm, defect density ≥ 10 1 5 cm − 3 ). That is 0.52 percentage points above the refined joint candidate and 8.9 points above the 18.52% reference. Surrogate‐proposed optima must therefore be validated within the originating simulator, and reported optima considered provisional pending such verification.