ABO3　type　perovskite　structures　are　promising　candidates　for　energy　storage　devices.　However,　their　recoverable　energy　storage　density　is　a　key　bottleneck　limiting　their　development.　It　is　noted　that　their　recoverable　energy　storage　density　can　tuned　by　introducing　different　elements.　Recurrent　neural　network　models　based　on　sequence　learning　are　proposed　to　predict　the　recoverable　energy　storage　density　when　different　elements　have　been　introduced　into　ABO3　type　perovskite　structures.　The　maximum　relative　errors　predicted　by　the　bidirectional　long　short　term　memory　network　models　with　Huble　loss　can　be　less　than　10　%.　This　means　that　the　proposed　method　can　be　used　to　rapidly　screen　the　recoverable　energy　storage　density.　Furthermore,　the　analysis　of　the　characteristic　importance　of　the　model　suggests　that　the　atoms　of　A　sites　and　B　sites　can　have　a　greater　contribution　to　the　recoverable　energy　density.　It　is　important　to　note　that　the　calculation　of　the　recoverable　energy　storage　density　is　a　challenging　and　computationally　intensive　task,　requiring　a　significant　amount　of　time.　Therefore,　its　prediction　is　a　difficult　problem.　This　paper　presents　a　promising　and　computationally　efficient　approach　to　this　problem.　©　2024　Elsevier　Ltd