Finding Oil With Neural Networks — By Hand
Running neural nets with a calculator to find oil, before the computers caught up.
Everybody talks about AI in the oil patch now like it showed up last year. It didn't. Some of us were running neural networks to hunt for oil a long time ago — we just didn't have a graphics card to do the heavy lifting. We had paper, a calculator, and a lot of stubbornness.
What "by hand" actually meant
A neural network, stripped down, isn't magic. It's a stack of weighted sums run through an activation function, nudged over and over until the output starts matching reality. Today a machine does billions of those nudges a second. Back then, I did them one at a time.
I'd take what I had — log data, readings, the patterns from wells that had already come in — and turn each signal into a number. Then I'd weight those inputs, add them up, push the result through an equation to decide whether the "neuron" fired, and compare the answer to what the ground had actually done. Wrong? Adjust the weights by hand, a little, in the direction that reduced the error. Then do it again. And again. What a modern model calls "training," I called a weekend.
Why anybody would do that
Because the question was worth it: is there oil under this spot, or not? Drilling a dry hole is one of the most expensive mistakes in the business. If I could stack the odds even a little — find the hidden relationship between the data I had and the formations that actually produced — that was real money and real risk, saved. The math was crude by today's standards, but the instinct was the same one every data scientist runs on now: let the patterns in the data you have tell you about the ground you haven't drilled yet.
The part that makes me laugh
Here's the funny part. I'd spend days grinding through calculations that a laptop now finishes before I've finished my coffee. The thing that used to take me a week of longhand arithmetic, a modern model does in a fraction of a second — and then does it ten thousand more times to check its work.
I'm not bitter about it. I think it's hilarious, and honestly a little vindicating. We weren't wrong about the idea. We were just early, and badly out-gunned on horsepower. The operators and engineers using machine learning to pick drilling locations today are standing on the same idea a handful of us were scratching out by hand years ago — we just had to do it the slow way.
What stayed the same
The computing power changed everything and nothing. Everything, because the scale and speed are unrecognizable. Nothing, because the job is still the same: respect the data, question your assumptions, and stay humble about what the ground will actually do. A faster model that forgets those things will drill a dry hole just as confidently as anyone ever did with a pencil.
I'll take the GPUs. But I earned my gray hairs doing it by hand — and I'd argue it taught me the math in a way no library import ever will.