I am trying to obtain the model equation (empirical formula) for predicting the optimum asphalt content for hotmix asphalt.

I have 12 variables with 5 nodes in one hidden layer (12-5-1 configuration).

I have used Levenberg-Marquardt backpropagation algorithm ('trainlm') in Matlab R2016a, with 'tansig' function for hidden layer and 'purelin' function for output layer to obtain weight and bias values.

Now I need to form a general equation of the output obtained from ANN, so that I may compare the result with dataset.

I tried to use the given weight and bias values by following equation:

y1 = tansig(W1*P+b1)

y2 = purelin(W2*y1+b2)

yet, the output I obtained is totally different with the reference data.

Optimum asphalt contents data have range from 6.1 to 6.3, meanwhile the predicted Optimum asphalt contents I obtained from ANN have range value -20's to 20's.

How did I get wrong??

The performance of the ANN output itself was very good, R = 0.9999968