Exploration geophysics is the discipline where ML methods entered exploration earliest and have penetrated deepest. The reason is structural: geophysical data is voluminous, multi-dimensional, and amenable to mathematical processing in ways that other exploration data are not. A regional airborne magnetic survey produces millions of measurements in a single flight; processing, inverting, and interpreting that volume of data benefits enormously from automated and semi-automated methods. The geophysics community has been integrating ML into standard workflows for fifteen years, well ahead of the rest of the exploration toolkit.
The current state in 2026 is that ML methods are now standard for several specific geophysics workflows: automated lineament and feature extraction, regularization of inversion solutions, neural-network-accelerated inversion for fast operational workflows, and joint integration of multiple geophysical datasets. The methods coexist with traditional physics-based inversion and interpretation rather than replacing them. The geophysics community has worked out, more than other exploration disciplines, where the ML/classical division of labor produces the best results.
Why Geophysics Was Early
The volume of geophysical data has driven methodology automation since the early days of large-scale airborne surveys. A modern airborne magnetic survey covers thousands of line-kilometers and produces measurements every fraction of a second along each line. Manual interpretation of this volume of data is structurally impossible; some form of automation is mandatory. The geophysics community has been developing and refining automated processing methods for decades, with ML adoption being a natural continuation of that trajectory rather than a disruption.
The data is also well structured for ML methods. Magnetic, gravity, EM, and IP measurements are spatial fields with well-understood physics. The transformations between data and interpretation — Fourier transforms, vector calculus operations, inverse problems — are mathematically standard. ML methods that operate on this structured data have clear physical interpretations and reproducible behavior, in contrast to ML applied to less structured exploration data.
And the academic-industry pipeline in geophysics has been strong. UBC's Geophysical Inversion Facility has been producing open-source inversion code (the GIF codes, now extended through the SimPEG framework) for two decades, with industry contributors including major mining companies, specialist consultancies, and government surveys. Mira Geoscience, Geosoft (now Seequent), and various smaller specialists have built commercial products around the same methodology foundation. The methodology evolution happens in the open and propagates into commercial workflows rapidly.
What Inversion Actually Does
Geophysical inversion is the mathematical process of going from surface measurements (e.g., the magnetic field at survey points) to a 3D model of subsurface properties (e.g., the magnetic susceptibility of subsurface rock). The classical inversion approach formulates this as an optimization problem: find the subsurface model that, when forward-modeled, produces measurements that match the observations, subject to regularization constraints that prevent overfitting to noise.
The methodology has been mature for decades. UBC-GIF's MAG3D, GRAV3D, DCIP3D, and EM3D codes implement the standard inversion methods for the major geophysical surveys. SimPEG, the open-source successor framework, generalizes the methodology and adds modern flexibility. Commercial products (VOXI, Geosoft's grid suite, Maxwell, Loki) implement variants of the same methodology with vendor-specific user interfaces and integrations.
The output of an inversion is a 3D model of the relevant physical property — susceptibility, density, conductivity, chargeability. This model is the geophysical contribution to the integrated exploration interpretation. It feeds into geological modeling, target identification, and drill planning, but it requires interpretation by a geophysicist or geologist with appropriate expertise to translate physical properties into geological meaning.
Where ML Augments Inversion
The strongest current application is in regularization and constraint enforcement. Classical inversion uses simple regularization (smoothness, sparsity) that doesn't always produce geologically reasonable models. ML methods can encode more sophisticated priors: trained on geological models from similar deposit types, an ML regularizer can bias the inversion toward solutions that match the structural and lithological characteristics typical of the target geology. The result is inversions that look more like real geology and require less interpretive cleanup downstream.
The second strong application is in joint inversion of multiple datasets. A magnetic survey alone produces a susceptibility model with substantial ambiguity. Combined with gravity (density), EM (conductivity), and surface geology, the joint solution is much more constrained. Classical joint inversion handles this through coupling terms in the optimization; ML methods can learn the cross-relationships between physical properties from training data and produce joint inversions that integrate the disparate data types more flexibly.
The third growing application is in accelerated inversion through neural network surrogates. A trained neural network that approximates the forward model — predicting what surface measurements a given subsurface model would produce — can run forward simulations in seconds rather than minutes, enabling far more iterations during inversion and far more thorough sensitivity analysis. The full inversion still uses the physics-based forward model for final solutions, but the iterative exploration that develops good initial models can use the surrogate.
Automated Lineament and Feature Extraction
A separate ML application area, distinct from inversion, is automated extraction of structural and geological features from processed geophysical data. CNNs trained on labeled lineament maps can identify fault traces, lithological contacts, and structural domain boundaries from magnetic and gravity processed grids more consistently than manual interpretation. The output is a candidate lineament map that the structural geologist reviews and refines.
The technology has reached maturity sufficient that several commercial products and research-grade systems are in routine use. The accuracy of the automated extraction depends on the quality of the input data and the relevance of the training set — automated lineament extraction trained on Precambrian shield magnetics may not perform well on basin geophysics — but for cases where the training data is well-matched, the productivity gains are substantial.
This is also the kind of work where the geophysicist's review remains essential. Automated lineaments include false positives from data artifacts, edge effects, and unrelated features. A 30-minute review by an experienced interpreter cleans up an automatic lineament extraction that took 30 seconds to produce, where the manual extraction would have taken 4-8 hours. The compression in productive time is meaningful.
Specific Tooling Worth Knowing
SimPEG remains the leading open-source framework for geophysical inversion. The methodology coverage is comprehensive (magnetic, gravity, DC resistivity, IP, EM, gravity gradiometry, joint inversion), the code is actively maintained, and the user community is large enough that learning resources are available. For teams committed to open-source workflows or doing methodology development, SimPEG is the natural foundation.
Geoscience ANALYST (Mira Geoscience) integrates geophysical inversion with implicit geological modeling and provides a unified visualization environment for geophysics-geology integration. The pricing model (free viewer, paid add-ons) is more accessible than fully commercial products and the integration with UBC-GIF and SimPEG is tight.
Geosoft Oasis montaj (Seequent) remains the dominant commercial platform for general-purpose geophysical processing, with VOXI Earth Modelling as the inversion add-on. The platform is widely used and well integrated with Leapfrog Geo. Cost is per-seat and meaningful.
For specialized methods — particularly time-domain EM, airborne EM, and magnetotellurics — vendor-specific software from the survey contractors (Aarhus Workbench, EMVision, WingLink, etc.) is often the right choice because the methodology is matched to the specific instrument and survey type.
The Realistic Capability Today
A modern exploration geophysics workflow on a moderately resourced junior project might look like: contract an airborne magnetic and EM survey, process the data through commercial software (Geosoft) with vendor support, run an inversion (SimPEG for cost-control, VOXI for commercial integration) to produce 3D physical property models, apply automated lineament extraction to the gridded magnetic data, integrate the geophysical interpretation with surface geology and geochemistry in a unified visualization (Geoscience ANALYST or Leapfrog Geo). The whole workflow can be executed by a small geophysics-geology team in weeks to a few months, where the same workflow would have taken half a year of consultant time a decade ago.
The cost compression is real but uneven. The survey acquisition cost has not dropped substantially — quality airborne work remains expensive. The processing and inversion cost has dropped significantly. The interpretation cost depends on whether the team has internal capability or contracts to specialists. For a junior with good in-house geophysics capability, the marginal cost of doing more sophisticated processing has dropped to near zero; for a junior that contracts everything, the cost structure is less dramatically improved.
What ML Doesn't Solve in Geophysics
The fundamental ambiguities of geophysical inversion don't go away with more sophisticated methods. A single dataset (magnetic alone, gravity alone) has infinite subsurface models that fit the data equally well; the inversion picks one according to its regularization, and that pick may not be the geologically correct one. Joint inversion and integration of geology reduces the ambiguity but doesn't eliminate it. The geophysicist's judgment about which solution to trust remains essential.
ML methods also struggle with novelty in the same way as other ML applications. A model trained on standard deposit-type geophysical signatures will identify more of the same; it will not flag a novel signature as "interesting" because it has no basis for that judgment. Discovery-focused geophysics work remains a human discipline where pattern recognition from the geophysicist's career experience is the limiting capability.
A Reasonable Starting Point
For a junior project with existing geophysical data that hasn't been re-processed with modern methods in several years, the highest-leverage first step is a re-inversion using current SimPEG or commercial inversion software, integrated with the project's current geological model. The re-inversion often surfaces structural and lithological features that the older processing missed, and the modern methodology gives the geophysicist much more interpretive flexibility.
For a project that's planning new airborne or ground surveys, the right investment is in upfront design quality: survey lines spaced for the deposit-scale features of interest, appropriate sensor configuration, careful processing protocols. The interpretive work that follows can be made faster with ML; the data acquisition decisions cannot be undone.
For consulting support on geophysical workflow integration or re-processing of historical data with modern methods, our free workflow audit covers exploration geophysics workflows, or contact us to discuss a pilot project.