The　purpose　of　this　study　was　to　systematically　evaluate　the　performance　advantages　of　the　Long　Short-Term　Memory　(LSTM)　in　predicting　the　axial　compressive　capacity　of　short　rectangular　concrete-filled　steel　tube　(RCFST)　columns　and　to　assess　its　effectiveness　in　interpreting　structural　mechanical　responses.　For　the　purpose　of　emphasizing　the　availability　of　the　LSTM　model,　a　controlled　experimental　design　was　implemented:　first,　the　Back　Propagation　Neural　Network　(BPNN)　was　established　as　a　benchmark　model,　and　subsequently,　the　prediction　results　of　both　models　were　compared　with　theoretical　calculations　derived　from　existing　formulas　across　multiple　indexes.　Through　a　quantitative　analysis　of　experimental　data　from　the　literature,　this　research　conducted　a　sensitivity　analysis　of　key　parameters　in　the　prediction　model　(such　as　cross　-sectional　area　of　steel　and　yield　stress　of　steel),　comparing　and　evaluating　the　influence　of　each　parameter　on　bearing　capacity.　The　results　indicate　that,　compared　to　the　BPNN　model,　the　LSTM　model　offers　significant　advantages,　demonstrating　higher　accuracy　and　reduced　discretization.　More　importantly,　the　weight　distribution　characteristics　of　the　LSTM　model　align　more　closely　with　the　structural　mechanical　mechanisms,　providing　a　valuable　reference　for　predicting　the　mechanical　properties　of　structures.