Seems you are doing time series prediction, and you must be using m previous values to pridict nth next value. to get inputs vectors from output you can build an inverse neural network having one input and m outputs. The accuracy of the network will be very low i guess.

Thanks sir,

But,

I am doing yeild maximization approach for a chemical reaction.

There are few( 3) process variable (input) and two response (output) such as

% yield of final product and final cost .

Now from the tested network I want to know the combination of input variables which would give maximum yield but lower cost ( both equal preference) .

This is known as optimization and used by desirability function (options) in other software package.

Even if I would get reply only for predixtion of set of input for single output , then also it serve my purpose great.

Thanking you again

inverse neural network or by suing weigh and bias , I think may help to solve it ?

Everything depends on the type of ANN you use. If the neural network is just feed forward and the network working in the real numbers you can theoretically switch the direction of all synapses and put to the network amount of energy to receive a input values. However, in case that neural network provides some normalisation, this will not work correctly.

plz go through the attached link 

http://www.iitbhu.ac.in/faculty/min/rajesh-rai/NMEICT-Slope/Pdf/15%20Introduction%20of%20softcomputing%20approach%20in%20slope%20stability.pdf

 ANNs can be used to predict the outcome of new independent input data

Maybe with some GAN (AC-GAN) architecture you can tackle the problem. Then you can say to the generator give an example for this type of output...