While　deep　learning　has　been　used　in　battery　computing　to　speed　up　the　search　for　new　cathode　materials,　the　majority　of　deep　learning　techniques　only　take　into　account　elemental　information　and　topological　information,　ignoring　the　significance　of　geometrical　information　and　global　information　for　electrode　average　voltage　prediction.　Multivalent　metal-ion　Battery　Voltage　Graph　Neural　Network　(MBVGNN)　proposed　in　present　work,　which　combines　global　information　and　geometric　information.　MBVGNN　achieves　43.98　%　improvement　over　reaction-based　GATGNN　and　47.06　%　over　TL-CGCNN　for　the　sodium-ion　battery　dataset.　Compared　with　the　best　available　models,　the　prediction　results　for　Ca,　Zn,　Al　and　Mg　electrode　materials　were　improved　by　15.50　%,　28.09　%,　44.74　%　and　18　%.　The　formation　energy　of　the　cathode　material　was　predicted　by　MBVGNN,　and　R2　was　0.969,　which　was　used　to　evaluate　the　thermodynamic　stability　of　the　cathode　material.　In　addition,　density　functional　theory　(DFT)　was　used　to　calculate　222　kinds　of　nickel　base　sodium-ion　battery　cathode　material＇s　average　voltage　for　high‑nickel　ternary　sodium-ion　batteries,　and　the　results　were　similar　with　the　predicted　results.　Screening　resulted　in　194　high-energy-density　ternary　sodium-ion　battery　cathode　materials.　Furthermore,　MBVGNN　can　reliably　predict　the　average　voltage　of　fluorine-substituted　layered　oxide　cathode　materials,　which　may　be　used　to　guide　the　experimental　synthesis　of　these　materials.　Based　on　the　accuracy　of　MBVGNN,　this　research　constructs　74,553　high-entropy　cathode　materials　that　include　layered　oxides　and　three　types　of　poly　anions.　Through　the　analysis　of　the　knowledge　relationship　obtained　from　the　prediction,　reference　guidance　for　16　material　types,　prioritized　elements　and　1,369,658　element　combinations　for　the　experimental　preparation　of　sodium-ion　battery　cathode　materials　is　provided.　This　research　employs　machine　learning　methods　to　demonstrate　that　the　current　preparation　of　high-voltage,　high-energy-density　cathode　materials　remains　necessary　through　multiple　doping　strategies.　©　2024　Elsevier　Ltd