Using a neural network as a data transfer function

Hello everyone,

I recently came up with the idea to use deep learning for one of my data evaluation tasks. I want to shortly describe the task so that you know if deep learning is even applicable.

We have a measurement system that measures light scattered by droplets in order to determine their size. The light signal is converted into a voltage signal which is recorded by a data acquisiton card. Droplets produce voltage peaks within the signal; these peaks are extracted. There are certain measurement errors that have a constant impact on the measurement so that the measured/extracted peaks usually look like an exponential decaying curve (lots of small voltage peaks, few high peaks). However, the corresponding droplet size distribution is a gaussian/weibull/gamma distribution. What I need is a transfer function that takes measurement errors into account; the problem is that these errors are not easily distinguishable. Therefore I was hoping to be able to train a network with calibration data: I have measured signal peak distributions for known and quite narrow droplet size distributions from a droplet generator. The trained network can afterwards be validated with measurements from another droplet generator that covers a broad range of droplet sizes. The input and output is in form of a histogram / bar graph, see figures below

Of course I already had a look into the deep network designer but was not sure how to design a network for my task. I thought about the sequence to sequence classification / regression but was not sure on how to define in and outputs. Basically what I need is a droplet size distribution output for a given signal peak distribution input. As far as I can tell, the sequence to sequence network is specifically designed for time-dependent data.

If anyone has an idea how to perform this task, please let me know.

Thank you,

Simon

Conversion of left distribution into right distribution?