Support Vector Regression (fitrsvm) with prior knowledge constraint

Hi there, I am using fitrsvm to estimate a regression model. The model fits data very well and overfitting is guaranteed since I use cross-validation. However, given the fact that the data is not rich enough to estimate the real physics model, I obtained the model which violates some physical constraints, i.e. partial derivative of the estimated function vs. independent variables must be positive/negative. I went through couple literature, such as https://hal.archives-ouvertes.fr/hal-00178619/document and see that this is a resolvable problem. Could anyone share with me how to incorporate these constraints into fitrsvm? Thanks a lot, Dave