Alloy　electrodes,　beneficial　from　excellent　stability,　are　considered　suitable　for　industrial　applications,　hence　exploring　alloy　catalysts　with　low　reaction　barriers　will　bring　innovative　scientific　understanding　and　enormous　economic　benefits.　Recently,　material　informatics　emerges　as　an　efficient　method　in　the　research　and　development　of　new　materials　through　diverse　candidates,　however,　collecting　a　large　amount　of　material　characterization　and　simulation　data　still　faces　numerous　difficulties.　To　tackle　this　issue,　combining　the　topological　structure　of　materials,　the　convolutional　neural　network　framework　developed　in　this　article　first　achieves　the　density　of　states　prediction　of　active　sites　on　the　alloy　surface,　based　on　which　the　adsorption　energy　of　different　reactants　is　obtained.　Benefited　by　electronic　structure,　this　model　exhibits　excellent　predictive　performance　with　a　mean　absolute　error　of　0.124　eV,　and　transferability　with　fast　convergence　under　dozens　transferred　data　to　complete　the　extension　for　high　entropy　alloys　and　reactants.　Based　on　this　massive　predictive　data,　high　entropy　alloy　catalysts　with　excellent　low　reaction　barrier　have　been　discovered,　and　several　catalytic　theories,　like　scaling　relations,　d-band　center　theory,　high-entropy　effects　and　synergistic　catalysis,　have　been　validated　and　improved.