This　paper　aims　to　predict　the　intake　characteristics　of　a　side-ported　Wankel　rotary　engine　implemented　by　a　novel　dual-layer　machine　learning　(ML)　structure.　The　Sobol　sequence　coupled　with　a　parametric　phase　control　model　is　performed　to　generate　samples　using　a　three-dimensional　dynamic　simulation　model.　Two　ML　structures　are　compared　to　determine　the　input　features,　in　which　the　input　features　of　structure　A　are　the　port　phases,　while　structure　B　also　contains　the　geometric　features.　Compared　with　Structure　A,　the　regression　and　generalization　abilities　of　Structure　B　with　different　ML　methods　are　generally　favorable.　Based　on　structure　B,　a　novel　dual-layer　structure　prediction　model　is　developed　using　the　Gaussian　process　regression　method　in　which　geometric　features　are　used　as　intermediate　variables.　The　results　show　that　for　the　prediction　of　intake　loss　and　volumetric　efficiency,　the　coefficients　of　determination　are　greater　than　0.99,　and　the　prediction　error　is　less　than　1%.　In　addition,　the　port　full　closing　timing　determines　the　intake　characteristics　compared　with　other　port　phases.　This　paper　provides　novel　solutions　for　predicting　intake　characteristics　to　guide　practical　optimization.(c)　2022　Elsevier　Masson　SAS.　All　rights　reserved.