Metasurfaces　find　a　wide　variety　of　applications　in　the　last　decades　due　to　their　powerful　ability　to　manipulate　electromagnetic　(EM)　waves.　Traditional　approaches　for　metasurface　design　require　massive　full-wave　EM　simulations　to　achieve　optimal　geometrical　parameter　values,　resulting　in　an　inefficient　design　process　of　metasurfaces.　In　this　article,　we　propose　a　physics-driven　machine-learning　(ML)　approach　incorporating　temporal　coupled　mode　theory　(CMT)　to　improve　the　design　efficiency　and　implement　an　intelligent　design　of　metasurfaces.　In　the　proposed　approach,　a　surrogate　model　(i.e.,　neuro-CMT　model)　is　developed　to　speed　up　the　prediction　of　EM　responses　of　metasurfaces.　A　three-stage　method　is　used　to　develop　the　neuro-CMT　model.　First,　we　perform　full-wave　EM　simulations　of　unit　cells　only　containing　single-　and　double-resonators　for　different　geometrical　design　parameter　values.　Second,　we　extract　the　single-　and　double-resonator　CMT　parameters　for　each　geometrical　parameter　value　by　fitting　the　corresponding　EM　responses　based　on　CMT　equations.　Third,　we　train　neural　networks　to　learn　the　relationships　between　the　CMT　parameters　and　geometrical　parameters　for　single-　and　double-resonator　systems,　respectively.　These　trained　neural　networks,　in　conjunction　with　the　multiresonator　CMT　equation,　become　an　efficient　tool　to　accurately　predict　the　EM　responses　of　any　arbitrary　coupled　multiresonator　systems.　The　proposed　neuro-CMT　model　can　be　further　utilized　for　metasurface　design　optimizations.　Two　metasurface　absorbers　are　given　as　examples　to　demonstrate　the　efficient　and　intelligent　advantages　of　our　proposed　approach.