Terahertz　radiation　is　essential　for　applications　such　as　optical　imaging,　data　transportation,　and　material　characterization.　However,　conventional　electronic　devices　struggle　to　generate　and　modulate　terahertz　radiation　effectively.　Metasurfaces　offer　a　promising　solution　by　manipulating　light　at　the　subwavelength　scale.　This　study　introduces　a　novel　approach　for　fabricating　subwavelength　optical　gratings　using　femtosecond　lasers　to　modulate　terahertz　reflection　and　transmission,　achieving　modulation　rates　ranging　from　0.3　%　to　99.9　%　across　wavelengths　of　100　to　200　mu　m.　We　developed　and　validated　two　machine　learning-based　forward　prediction　models　with　R2　values　of　0.999　and　0.982,　respectively,　to　predict　the　structure　and　optical　properties　of　laser-　fabricated　gratings.　Additionally,　two　inverse　design　models　were　constructed,　enabling　precise　determination　of　optimal　design　structures　and　laser　processing　parameters　for　desired　optical　properties,　with　R2　values　of　0.989　for　two-structure　models　and　0.878　for　three-structure　models.　The　lower　accuracy　of　the　three-structure　model　reflects　its　inherent　complexity.　Experimental　validation　confirms　the　effectiveness　of　these　neural　network-　based　models　in　advancing　the　manufacturability　and　application　of　terahertz　metasurfaces.