Where This Episode Starts

We begin with a fundamental limitation in Earth science:
we cannot directly measure most of what lies below the surface.

Geophysics addresses this problem by combining physics, geology, and mathematics to infer subsurface properties—such as electrical conductivity, density, or seismic velocity—using measurements collected at or above Earth’s surface. These measurements are often gathered under extreme conditions, from helicopters or on foot, across remote terrain.

The challenge is not collecting the data—it’s interpreting it.

---

What Is a Geophysical Inversion?

A geophysical inversion is a mathematical process that works backward from observations to estimate the physical properties that must exist underground to produce those measurements.

Because Earth cannot be sampled directly at depth, inversions attempt to solve an ill-posed problem:
there are far more unknowns than knowns, and many different models can fit the same data equally well.

This creates an entire space of possible solutions rather than a single, definitive answer.

---

The Problem of Ambiguity

Traditional geophysical inversions rely on assumptions—often favoring smooth, continuous models—to reduce the infinite number of possible solutions.

These assumptions, known as regularization, help stabilize the math but can blur real geological boundaries. In nature, transitions between rock types are often sharp, not smooth, and standard approaches can fail to capture this structure.

This is where newer tools begin to matter.

---

Enter Machine Learning

Rather than replacing geophysics, machine learning in this episode is framed as a tool for guiding interpretation, not automating it.

Jonathan Kataj’s work focuses on using image segmentation methods, commonly applied in computer vision, to identify boundaries and structure within geophysical models. These methods allow the inversion process to incorporate spatial information—such as orientation, layering, and discontinuities—more explicitly.

The goal is not to guess the answer, but to better constrain the range of plausible solutions.

---

From Smooth Models to Structured Earth

A key concept explored in the episode is how prior geological knowledge can be encoded mathematically.

Using approaches such as Gaussian mixture models and Markov random fields, it becomes possible to represent multiple rock units, each with distinct physical properties, while also accounting for how those units interact spatially.

This allows inversions to move beyond overly smooth images toward models that better reflect real geological structure.

---

Data Collection in the Real World

Behind every inversion lies data collected under challenging conditions.

The episode highlights what fieldwork actually looks like:
kilometers of wire dragged across difficult terrain, long days in remote regions, equipment failures far from help, and constant problem-solving. These realities shape both the quantity and quality of the data—and ultimately influence what can be inferred from it.

Machine learning does not remove these constraints; it helps scientists make better use of imperfect information.

---

From Industry to PhD Research

Jonathan Kataj’s path into this research was not linear.

The conversation traces a journey through engineering, industry data acquisition, software development, and eventually doctoral research at the University of British Columbia. Along the way, exposure to real-world limitations in existing tools motivated the development of new methods.

This episode highlights how scientific progress often comes from lived experience with broken or insufficient systems.

---

Publishing the Science

The episode also walks through the research publication process, including Jonathan’s first peer-reviewed paper, accepted in Geophysics, a journal of the Society of Exploration Geophysicists.

Rather than focusing on results alone, the discussion emphasizes how methods papers contribute to the broader scientific toolbox—allowing others to apply, adapt, and build on the work.

---

Why This Matters

Understanding Earth’s subsurface is critical for studying natural hazards, groundwater, mineral resources, and planetary processes.

This episode illustrates how combining physics-based models, field data, and machine learning can reduce uncertainty without abandoning scientific rigor. Rather than treating AI as a black box, it is presented here as a way to encode geological intuition into mathematical frameworks.

---

Episode Format

• Long-form scientific conversation

• Geophysics and Earth imaging

• Machine learning in applied science

• Research methods and scientific publishing

Topics & Keywords

geophysics, machine learning in science, geophysical inversion, artificial intelligence, subsurface imaging, image segmentation, Earth structure, scientific computing, applied mathematics, research methods

---

Enjoyed this episode?

Subscribe to Whimsical Wavelengths for deep dives into how science actually works—from the familiar to the most extreme environments in the universe.