The Equation That Sees Underground: Why Seismic Imaging Is an Inverse Problem

We never actually see underground.

No matter how advanced our satellites are, no matter how sharp our maps look, the subsurface remains opaque. What lies beneath our feet — layers of rock, trapped fluids, fractures, faults — is hidden from direct observation.

And yet, oil is drilled. Aquifers are mapped. Earthquakes are analyzed. How?

We don’t see the Earth. We listen to it.

And the language it speaks is a Partial Differential Equation.

The Earth Reveals Itself Through Waves

When energy is released at the surface a seismic source, an explosion, a vibration it travels through the Earth as waves. These waves reflect, refract, and scatter depending on the properties of the rocks they pass through.

What we record at the surface is not the Earth itself, but the response of the Earth.

Mathematically, this process is governed by the wave equation.

In its simplest acoustic form:

(1 / v²(x)) · ∂²u/∂t² = ∇²u + s(x,t)

Where:

• u(x,t) is the wavefield
• v(x) is the subsurface velocity model
• s(x,t) is the source term
• ∇² captures spatial curvature of the wave

This equation is not optional. Every seismic wave obeys it.

Forward Problems: When Physics Is Kind

If I give you:

• the velocity model v(x)v(x)v(x)
• the source s(x,t)s(x,t)s(x,t)
• boundary and initial conditions

You can simulate how waves propagate through the Earth.

This is called a forward problem.

It is:

• well-posed
• deterministic
• stable

Classical numerical solvers handle this well. Even modern machine learning can approximate it.

But forward problems are not where geology gets difficult.

The Real Problem: Inversion

In the real world, we face the opposite situation.

We don’t know the velocity model.

All we have are:

• wave recordings at the surface
• noisy, incomplete, indirect measurements

And from those, we want to infer the subsurface.

This is an inverse problem.

Formally:

Given observations of u(x,t)u(x,t)u(x,t), infer the parameters v(x)v(x)v(x) that produced them.

This is where everything breaks.

Why Inverse Problems Are Fundamentally Hard

Inverse problems in Earth science are:

• Ill-posed Multiple subsurface models can explain the same data.
• Unstable Small noise in measurements can cause large changes in solutions.
• Underdetermined We never have enough data to uniquely constrain the Earth.

This is not a software issue. It is a mathematical reality.

No amount of data alone fixes this.

Why Data-Driven AI Struggles Underground

Most modern AI systems are trained to minimize data loss:

L_data = || u_pred — u_obs ||²

This works well when:

• patterns are directly observable
• the problem is interpolation-heavy

But seismic imaging is not interpolation. It is inference under physical constraints.

A neural network can fit seismic data perfectly while producing a subsurface model that violates the wave equation entirely.

It may look right. It may plot nicely. But it is physically wrong.

The Earth does not care about accuracy metrics. It only obeys physics.

Inverse Problems Are Optimization Problems with Constraints

At its core, seismic imaging is a PDE-constrained optimization problem.

We are not just minimizing mismatch with data. We are minimizing it subject to the wave equation.

Conceptually:

min || u_sim(v) — u_obs ||² such that N[u, v] = 0

Where:

• min is the minimization over v(x)
• N represents the governing PDE.

This constraint is everything.

Without it, inversion collapses into guesswork.

Where Scientific Machine Learning Enters

This is where Scientific Machine Learning (SciML) — and specifically Physics-Informed Neural Networks (PINNs) — becomes interesting.

PINNs do not replace physics. They embed it.

Instead of training a network only on data, we penalize violations of the governing equation:

L_physics = || N[u_θ] ||²

The model is not rewarded for being accurate alone it is rewarded for being physically consistent.

This is a subtle but profound shift.

What PINNs Can — and Cannot — Do in Seismology

Let’s be honest :

PINNs are not a silver bullet for full-scale seismic inversion.

• High-frequency wavefields are hard
• Training can be unstable
• Scaling is still an open challenge

But conceptually, PINNs expose something crucial:

Inverse problems cannot be solved by data alone. They require structure.

And physics provides that structure.

Hybrid approaches combining classical solvers, adjoint methods, and SciML are likely the real future.

Why This Matters (Especially Here in Africa)

In Africa, subsurface uncertainty affects:

• groundwater exploration
• geothermal energy
• mineral resources
• seismic risk assessment

These are not abstract problems.

They are inverse problems governed by PDEs.

The question is not whether we use AI. The question is what kind.

The Quiet Truth About Seeing Underground

The Earth is not hiding its behavior.

It writes it down clearly in equations.

The challenge of geophysics is not prediction. It is interpretation under constraint.

If we want machines to help us see underground, we must first teach them to respect the wave equation.

Because underground, physics is not a suggestion. It is the law.