Recently,　sharded-blockchain　has　attracted　more　and　more　attention.　Its　inherited　immutability,　decentralization,　and　promoted　scalability　effectively　address　the　trust　issue　of　the　data　sharing　in　the　Internet　of　Things　(IoT).　Nevertheless,　the　traditional　random　allocation　between　validator　groups　and　transaction　pools　ignores　the　differences　of　shards,　which　reduces　the　overall　system　performance　due　to　the　unbalance　between　computing　capacity　and　transaction　load.　To　solve　this　problem,　a　load　balance　optimization　framework　for　sharded-blockchain　enabled　IoT　is　proposed,　where　the　allocation　between　the　validator　groups　and　transaction　pools　is　implemented　reasonably　by　deep　reinforcement　learning　(DRL).　Specifically,　based　on　the　theoretical　analysis　of　the　intra-shard　consensus　and　the　final　system　consensus,　the　optimization　of　system　performance　is　formed　as　a　Markov　decision　process　(MDP),　and　the　allocation　of　the　transaction　pools,　the　block　size,　and　the　block　interval　are　jointly　trained　in　the　DRL　agent.　The　simulation　results　show　that　the　proposed　scheme　improves　the　scalability　of　the　sharded　blockchain　system　for　IoT.　Copyright　©　by　HIGH　TECHNOLOGY　LETTERS　PRESS.