Neural　network　based　techniques　are　useful　in　representing　the　behavior　of　electronic　devices　for　nonlinear　circuit　simulation.　As　new　technologies　evolve,　new　behavior　need　to　be　captured　for　modeling　and　simulation.　Often,　measurement　data　from　the　devices　are　used　to　train　a　neural　model　to　capture　these　behaviors.　However,　during　nonlinear　simulation,　the　nonlinear　equations　derived　from　the　electronic　circuits　need　to　be　solved　iteratively,　and　the　iterative　variables　may　explore　outside　the　training　data　range,　potentially　causing　convergence　failure.　This　paper　addresses　such　problem　by　describing　an　extrapolation　technique　that　extrapolate　the　information　smoothly　beyond　the　training　data　range.　To　guarantee　convergence　of　nonlinear　simulation,　the　extrapolated　model　is　made　maximumly　smooth　while　satisfying　diverse　tendencies　of　the　model　in　multi-dimensional　parameter　space.　The　technique　is　demonstrated　by　a　high-electron-mobility　transistor　modeling　example　and　its　use　in　harmonic-balance　simulation,　showing　better　convergence　of　nonlinear　circuit　simulation　using　extrapolated　neural　models　over　existing　neural　modeling　methods.　©　2019　IEEE.