Abstract Dion–Jacobson two-dimensional hybrid organic–inorganic perovskites are promising ferroelectric materials, combining chemically tunable polarity enabled by their organic spacers with enhanced environmental stability. However, rational discovery remains difficult owing to uncertain phase formability and scarce, biased ferroelectric data. Here we show a staged, physically guided synthesis-by-design workflow that integrates learning from scarce ferroelectric data with structure-aware screening. Starting from 11 known ferroelectric compounds in this family, symbolic regression defines a descriptor-bounded ferroelectric-consistent region. Similarity-guided expansion and chemical filtering reduce 83,924 molecules to 1012 candidate divalent spacers, while machine-learned interatomic-potential relaxation and connectivity analysis identify 338 candidates retaining Dion–Jacobson connectivity. Sequential prioritization by connectivity retention, chemical similarity, energetic refinement and synthetic accessibility selects s −2-fluorobutane-1,4-diamine, leading to the synthesis of a PbI 4 -based Dion–Jacobson perovskite with experimentally confirmed ferroelectric switching. This prospective proof of concept shows how sparse property knowledge can be coupled with structural screening to prioritize organic spacer candidates for experimental synthesis.