You would first need to do a screening study with your 3 factors in jmp, I would suggest a 2X3 (8 runs) with centre points to account for curvature rather than a 3X3 (27 runs). Use the screening study to remove factors that are not significant. Check for the validity of your model by use of graphical methods (residuals and a normal plot) and numerical methods (ANOVA and lack of fit). Once the model is shown to be sound, fit in a linear model based on your significant factors - this linear model can be used to do a method of steepest ascent if your initial screening didn't identify the region of maximum yield - After this, augment your initial screening study with axial points and carry out a response surface study using a central composite design (CCD) or a Box-Behnken design (depends on your factor ranges). The response surface method (RSM) will give you your optimum values by analyzing the prediction profilers, setting desirability to maximize yield and reading response surface curves or contour profilers. Finally, fit in the second-order model.