This　study　proposes　a　three-level　meta-frontier　framework　enhanced　with　machine　learning-driven　projection　methods　to　address　the　dual　heterogeneity　in　carbon　emission　efficiency　analysis　arising　from　regional　disparities　and　industrial　diversification.　Methodologically,　we　introduce　two　novel　projection　combinations-＂exogenous-exogenous-accumulation　(E-E-A)　and　exogenous-exogenous-consistent　(E-E-C)＂-to　resolve　the　inconsistency　of　technology　gap　ratios　(TGRs　＞　1)　in　traditional　nonradial　directional　distance　function　(DDF)　models.　Reinforcement　learning　(RL)　optimizes　dynamic　direction　vectors,　whereas　graph　neural　networks　(GNNs)　encode　spatial　interdependencies　to　constrain　the　TGR　within　[0,　1].　Empirical　analysis　of　60　countries　reveals　that　(1)　E-E-C　eliminates　the　TGR　overestimation　by　12-18%　in　energy-intensive　sectors　(e.g.,　reducing　Asia＇s　secondary　industry　TGR1　from　1.160　to　1.000);　(2)　industrial　heterogeneity　dominates　inefficiency　in　Asia　(IHI　=　0.207),　whereas　management　gaps　drive　global　secondary　sector　inefficiency　(MI　=　0.678);　and　(3)　policy　simulations　advocate　for　decentralized　renewables　in　Africa,　fiscal　incentives　for　Asian　coal　retrofits,　and　expanded　EU　carbon　border　taxes.　Computational　enhancements　via　Apache　Spark　achieve　a　58%　runtime　reduction.　The　framework　advances　environmental　efficiency　analysis　by　integrating　machine　learning　with　meta-frontier　theory,　offering　both　methodological　rigor　(via　regularization　and　GNN　constraints)　and　actionable　decarbonization　pathways.　Limitations　include　static　heterogeneity　assumptions　and　data　granularity　gaps,　prompting　the　future　integration　of　IoT-enabled　dynamic　models.