Machine learning based fast forward modelling of ground-based time-domain electromagnetic data

Thue Sylvester Bording, Muhammad Rizwan Asif*, Adrian S. Barfod, Jakob Juul Larsen, Bo Zhang, Denys James Grombacher, Anders Vest Christiansen, Kim Wann Engebretsen, Jesper Bjergsted Pedersen, Pradip Kumar Maurya, Esben Auken

*Corresponding author af dette arbejde

Publikation: Bidrag til tidsskrift/Konferencebidrag i tidsskrift /Bidrag til avisTidsskriftartikelForskningpeer review


Inversion of large-scale time-domain electromagnetic surveys are computationally expensive and time consuming. Deterministic or probabilistic inversion schemes usually require calculations of forward responses, and often thousands to millions of forward responses are computed. We propose a machine learning based forward modelling approach as a computationally feasible alternative to approximate numerical forward modelling where a neural network is employed to model the relationship between the resistivity models and corresponding forward responses. For training of the neural network, we generated forward responses using conventional numerical algorithm for 93,500 resistivity models derived from different surveys conducted in Denmark representing typical resistivities of sedimentary geological layers. The input resistivity models and the network target outputs, i.e. forward responses, are scaled using a novel normalization strategy to ensure each gate is equally prioritized. The performance of the network is evaluated on two test datasets consisting of 8942 resistivity models by comparing the forward responses generated by the neural network and the conventional algorithm. We also measure the performance for the time derivatives of forward responses, i.e. dB/dt, by incorporating a system response. The results show that the proposed strategy is at least 13 times faster than commonly used accurate modelling methods and achieves an accuracy of 98% within 3% relative error, which is comparable to data uncertainty. Additional experiments on surveys from two other continents show that the results generalize in similar geological settings. Thus, under certain geological constraints, the proposed methodology may be incorporated into the pre-existing inversion structures, allowing for significantly faster inversion of large datasets.

TidsskriftJournal of Applied Geophysics
Antal sider9
StatusUdgivet - apr. 2021


Dyk ned i forskningsemnerne om 'Machine learning based fast forward modelling of ground-based time-domain electromagnetic data'. Sammen danner de et unikt fingeraftryk.