With　the　continuous　increase　in　modern　data　scale,　there　is　a　higher　demand　for　processor　computing　power　for　time-consuming　and　computationally　intensive　data　mining　problems.　High　performance　computing　(HPC)　can　enable　faster　processing　of　large　datasets　and　complex　algorithms.　Data　prefetching　is　an　effective　technique　used　in　HPC.　It　can　hide　memory　delay　and　boost　execution　performance　by　obtaining　data　before　the　program　needs　it.　Traditional　data　prefetchers　show　poor　performance　on　programs　with　irregular　access　patterns　due　to　the　complex　access　patterns　in　its　workload　is　difficult　to　predict,　and　may　lead　to　serious　cache　miss.　Applying　machine　learning　algorithm　into　data　prefetching　policy　for　access　prediction　can　effectively　improve　system　performance.　In　this　paper,　we　propose　a　machine　learning-based　prefetcher　called　Adaptive　Multilayer　Perceptron　Prefetcher　(AMPP),　combined　with　a　cosine　annealing　learning　rate　scheduler.　AMPP　regards　prefetching　as　a　classification　problem,　it　analyzes　the　given　instruction　pointer　using　a　multilayer　perceptron　model　and　can　use　the　learning　rate　scheduler　to　optimize　the　model　during　the　training　process.　We　apply　AMPP　on　last　level　cache　and　use　ChampSim　simulator　to　conduct　simulation　experiments　based　on　GAP　and　SPEC　CPU　benchmark　suits.　Our　experiments　demonstrate　that　AMPP　provides　average　IPC　improvement　of　35.58%,　MPKI　improvement　of　46.08%　and　coverage　of　51.32%　compared　to　no-prefetching　baseline.　And　it　performs　especially　well　on　benchmarks　with　irregular　data　access　patterns.　©　2023　IEEE.