I would like to create a machine learnig framework that could predict the 3D heat distribution of a room(of size 120x120x120) , given multiple parameters(position of the heater, orientation, power of the heater, etc).
For the moment, I will keep fixed all the parameters except the orientation, which is represented as two angles(let's call them theta and phi). I will fix theta=0 and only vary phi. If that works I will extend the approach for more parameters.
In cosequence , I would like to develop a Neural Net that takes phi ( from -50° to 90°) and predicts the heatmap(numbers from 0 to 1e7)
I am currently trying with a single Deconvolutional network, but It seems hard to learn the correlation between input and output. Even I try to overfit a single data and it doesn't work .
I have already experience with regression using U-Net and MLP, but this case is quiete different because I'm starting from a single value that is not related to an image feature, and trying to generate an image.
Any suggestions to face this problem? Maybe the use of GAN's could help me in someway ? Maybe the use of a discriminator could help me to achieve this better. Another approach that you could me recommend?