AI That Learns Physics Can Look Right But Be Wrong

Imagine teaching an AI to copy a rollercoaster by showing it thousands of photos. It might draw the track perfectly — every hill and dip in the right place. But if you then asked it how steep the track is at any given point, or how sharply it curves, it could be badly wrong. That's the finding of a new paper on "physics-informed neural networks" — AI systems trained not just on data, but on the actual equations that govern how things move and change.

The researcher trained these AI models only on ordinary values of two simple functions, sin(x) and exp(x) — the wavy and the fast-growing curves. Then he checked, separately, whether the AI could correctly calculate the rate of change (first derivative) and the rate of that change (second derivative). The answer was no. The AI's drawings looked correct to the eye, but its math was off, sometimes substantially — and the errors were worst near the sharpest bends and at the edges of the range.

To make sure this wasn't a fluke, he ran the tests again with more training points, different internal settings, and two different ways of measuring error. The problem held up. That's important because these models are used in real engineering: simulating airflow over a wing, heat moving through a chip, groundwater flowing underground, or weather patterns. In all of those, the derivative is often the thing you actually care about — it's what tells you force, stress, flow rate, or risk.

The paper's real contribution is a simple safety habit. It proposes a check-up routine that separates "does the output look right?" from "is the underlying physics reliable?" Think of it as a car that looks flawless in the showroom but needs the brakes tested on a track. The takeaway for anyone using AI for science or engineering: a pretty graph is not proof. Ask for the derivative test too, and be extra suspicious near the edges.