2026-08-12 –, Room 6
This talk discusses two complementary directions in scientific machine learning for geophysics. The first uses DeepONet surrogates to accelerate magnetotelluric forward modelling and transdimensional probabilistic inversion, making uncertainty analysis more practical. The second uses implicit neural representations for three-dimensional gravity inversion, where the subsurface model is learned under physics-based machine learning.
The first direction concerns neural networks trained as surrogates for geophysical forward problems. I will present a workflow in which a DeepONet is trained to emulate one-dimensional magnetotelluric responses and is then used within a transdimensional probabilistic inversion scheme. By replacing repeated evaluations of the full forward solver with a learned surrogate, the method makes large-scale posterior exploration and uncertainty analysis substantially more practical. The example shows how surrogate modelling can reduce computational cost while preserving the level of accuracy needed for probabilistic inversion.
The second direction concerns the use of neural networks as the inversion model itself. In this setting, the subsurface is represented as an implicit neural field constrained by the governing physics rather than by a fixed voxel parameterisation. I will illustrate this with three-dimensional gravity inversion, where implicit neural representations recover both smooth variations and sharp contrasts while reducing the reliance on explicit depth weighting. The gravity examples show that neural-field parameterisations can produce geologically plausible models while remaining compact and flexible.
Pankaj K Mishra is a Senior Scientist (Geophysics) at Geological Survey of Finland.
For more info visit: https://pankajkmishra.github.io/