How to remove complex numbers from feval

When I run this code to evaluate the equations (using simple feval functions) I get complex answers resulting in the plot shown. How does one handle this in the code to get a single line fit that I need and get rid of the complex solutions?

The 2 column table being read in is ejust xperimental data from uni thats been cleaned and Ive tried using statements with real(x) etc that doesnt work.

Many thanks for the help

GSI = 100;
D = 0;
Sig3 = TeoFQI.Sig3;
Sig1 = TeoFQI.Sig1;
HoekBrown = @(b,Sig3)(Sig3+b(1)*((b(2)*exp((GSI-100)/(28-14*D))*(Sig3/b(1))+ exp((GSI-100)/(9-3*D)))).^(0.5+1/6*(exp(-GSI/15)-exp(-20/3))));
opts = statset('nlinfit');
beta0 = [100;7];                              
beta = nlinfit(Sig3,Sig1,HoekBrown,beta0,opts)
S3min = min(Sig3)
S3max = max(Sig3)
figure()
plot(Sig3,Sig1,'o')
hold on
S3 = linspace(-6,S3max+4,100)
HB=@(Sig3x) Sig3x+beta(1)*((beta(2)*exp((GSI-100)./(28-14*D)).*(Sig3x/beta(1))+ exp((GSI-100)/(9-3*D)))).^(0.5+1/6.*(exp(-GSI/15)-exp(-20./3)));
y = feval(HB,S3)
plot(S3,real(y))
hold off
mb = beta(2).*exp((GSI-100)./(28-14*D));
s = exp((GSI-100)/(9-3*D));
a= 0.5+1/6*(exp(-GSI/15)-exp(-20/3));
SigciHold = beta(1)
SigN = @(Sig1x,Sig3x) ((Sig1x+Sig3x)/2 - (Sig1x-Sig3x)/2*(((1+a.*mb.*(mb*Sig3x./SigciHold+s).^(a-1))-1)/((1+a.*mb.*(mb*Sig3x./SigciHold+s).^(a-1))+1)));
SigN_Evaluated = feval(SigN,y,S3)
Tau = @(Sig1x,Sig3x) ((Sig1x-Sig3x).*(sqrt(1+a.*mb.*(mb.*Sig3./SigciHold+s).^(a-1)))./((1+a.*mb.*(mb.*Sig3./SigciHold+s).^(a-1))+1));
Tau_Evaluated = feval(Tau,y,S3)
figure()
plot(SigN_Evaluated,Tau_Evaluated,'o')

[Walter Roberson]

It sounds as if you need a technique that supports bounds.

I wonder if https://www.mathworks.com/help/matlab/ref/lsqnonneg.html could be used? Or something from the Curvefitting Toolbox, as that supports lower and upper bounds.

[Ken Mercer]

Ive tried changing the bounds but no difference still results in a complex solution

[Walter Roberson]

nlinfit() does not appear to me to support bounds parameters.

It would help us to test if you supplied the data.

[Ken Mercer]

Walter many thanks you were corrent it is a bounds issue. Now resolved.