Can machine learning, AI, analytics and the IoT add to AEM processing and interpretation?

Please join us for a special seminar next Tuesday March 26 at 3:30 pm in Biology 106 (note location change) with KEGS speaker Jim McNae, RMIT University:

Can machine learning, AI, analytics and the IoT add to AEM processing and interpretation?

Electromagnetic (EM) data is being collected at ever higher streaming rates, with airborne AEM data sampling rates approaching 1 MHz in some systems, and ground penetrating radar (GPR) sampling approaching 1 GHz. Six hours of data acquisition with BIPTEM, a 24-bit, 12 channel AEM system (6 B field sensor, 3 dB/dt, 3 rotation rate), each channel sampled at 156250 Hz will deliver over 150 GB of data.

To extract “useful” physical property information from this mountain of data, and thereby infer useful geology, there are many options. The historically most useful process combines physical insight to infer conductivity from the observed response with statistical methods to improve signal/noise. For example, EM data from a controlled source survey are presented as profiles or inverted with logarithmic time spacing, sensible for EM diffusion. Reduced noise has come from e.g. binary stacking, and recognition that sferic source energy is non-stationary and can usefully be “pruned” from the data before stacking. Subsets of the acquired data are then selected, modelled and inverted based on simple models or a-priori assumptions. Questions remain as to whether commonly applied a-priori assumptions are reasonable, whether all the useful implications of the data, such as induced polarization (AIP) and superparamagnetic (ASPM) effects have been extracted, and whether sferics, powerline signals and VLF can provide complementary conductivity information in a controlled source survey.

Some recent papers in computer science related to the “Internet of Things” (IoT) have presented examples of detecting anomalies in IoT time series using deep (machine) learning. These examples have some elements in common with AEM data processing and interpretation, and may in the future lead to automated QC and first pass physical property prediction and ultimately geological interpretation. However, implementing and setting up such processes will still be a very significant challenge for geoscientists. I therefor suspect that the Australian mineral explorer that last year made its geologists and geophysicists redundant, and advertised for “data mining specialists” will be too far ahead of its time.