Background - often we hear people question neural network (NN) as a black box. Recall in geoscience, we often validate an earth model using seismic-well calibration at selected well locations. We prefer a white box.

Action - seek black-to-white box pathway, in July 2018, AAPG hosted the inaugural hackathon.

https://www.aapg.org/publ...ntists-and-engineers

Follow up - experiment with neural network-driven decline curve analysis (DCA).

Input: production decline curves

Output: physics-strong ML experiments:

Baseline - deterministic model-based approach. Automate best fit parameters to the Modified Arp’s equation with SciPy (Python scientific library for numerical optimization, serving up all the goodies like least square, conjugate gradient etc.)

ML simple - NN single fully-connected layer with varying number of neurons.

ML deep - recursive neural network (RNN), specifically long-short term memory (LSTM), the most powerful ML model for time series.

Learning with machine, it is important to have ML-free reference. Say play with the Arp’s parameters. (Warning - it can be highly addictive and easily gobbles up 20-min per DCA. One Saturday morning, I was so absorbed and missed the good part of a Hotspur EPL game.)

Perspective on physics-strong ML - with simple NN-DCA, we can treat the weights of single-layer NN as filter coefficients, as in digital signal processing, or predictive decon in seismic data processing. We then tap the full gamut of Fourier analysis and discern what NN really does. Even more salivating, we unlock the corpus of knowledge in wavelet analysis from geophysics and may come up with better NN design.

LSTM-DCA is a little more complicated. There is more than one layer, visualize multi-dimensional filter? Nevertheless, it may be worthy of further investigation and research. As food for thought -
http://ataspinar.com/2018...in-machine-learning/