That looks like a version of the Lognormal Distribution. It would likely be best to code the ‘M’ result as a function of ‘H’, and then use one of the optimisation routines to solve for the ‘D_m’ as a parameter.
Code it, then first experiment with lsqcurvefit or another routine. If that approach has problems, consider using a Global Optimization Toolbox function such as ga to see if it can be estimated.
