Per-　and　polyfluoroalkyl　substances　(PFAS)　have　emerged　as　global　environmental　contaminants　owing　to　their　persistence,　bioaccumulation　potential,　and　toxicity.　Among　various　remediation　approaches,　foam　separation　(FS)　has　attracted　increasing　attention　for　its　simplicity,　cost-effectiveness,　and　low　energy　demand.　However,　the　PFAS　removal　efficiency　of　FS　is　highly　sensitive　to　pollutant　properties,　operational　parameters,　and　environmental　conditions,　which　vary　substantially　across　studies　and　pose　challenges　for　process　optimization.　To　address　this　issue,　six　machine　learning　(ML)　models　were　evaluated,　among　which　extreme　gradient　boosting　(XGB)　exhibited　the　highest　predictive　performance　(R2).　Based　on　this,　an　improved　model,　PFAS-XGB,　was　developed　to　streamline　input　variables　and　enhance　prediction　efficiency.　Model　interpretation　identified　aeration　time　and　PFAS　molecular　weight　as　the　most　critical　determinants　of　removal　efficiency.　Causal　contribution　analysis　indicated　that　operational　conditions　accounted　for　56.8　%　of　the　influence,　followed　by　PFAS　properties　(21.9　%),　surfactant　characteristics　(16.8　%),　and　metal　activators　(4.5　%).　These　results　underscore　the　predominant　role　of　operational　parameters　while　revealing　the　previously　underestimated　impact　of　surfactants　and　metal　activators.　Overall,　this　study　establishes　a　data-driven　framework　for　optimizing　FS　performance　and　offers　actionable　insights　for　enhancing　PFAS　remediation　strategies.　©　2025　Elsevier　Inc.