When AI Learns Physics…
Most industrial AI can notice patterns but cannot explain the physical story underneath them. Physics-informed neural networks change what the field can ask of a model.
These are the companion notes. The full essay lives on The Turing Pilgrim, my Substack publication.
Why this matters
A dashboard flashing “abnormal” is not a decision. A compressor does not fail because a model score crossed 0.73 — something physical happened, and in the field, cycles are spent working out what. Most industrial AI today can flag the pattern but cannot tell the physical story underneath it, and the story determines the action.
Physics-informed neural networks are trained with two kinds of discipline: they learn from sensor data, and they are penalized when their predictions violate known physics — conservation of mass and energy, fluid flow, heat transfer, boundary conditions. That backbone matters because industrial data is never as clean as the demo. Physics narrows the range of nonsense, and in the field, narrowing the range of nonsense is already a huge win.
The deeper shift is inference. Sensors are sparse; the physical world is continuous. A physics-aware model can estimate the internal state of an asset you cannot directly observe — which is the difference between “failure risk high” and “this pressure behavior is consistent with early liquid loading.”
Key takeaways
- Pattern machines miss causes. Bearing wear, misalignment, and sensor drift can produce the same ugly signals but demand different responses. The “why” determines the action.
- PINNs add a physics penalty to training. The model is wrong if it misses the data, and wrong again if it violates known physics.
- CFD becomes interactive. Fast physics-aware surrogates let the field ask what-if questions without a multi-day engineering study.
- Maintenance becomes causal. Not “risk high” but “consistent with increasing imbalance and bearing degradation, not a cooling issue.”
- Nobody buys PINN software. The product is answers: what is happening, why, how urgent, and what happens if you wait.
- Start narrow. A bounded asset class, a valuable failure mode, sparse sensors, expensive decisions. Then prove it.
Who should read it
Product and operations leaders responsible for predictive maintenance, thermal management, or flow assurance — anyone deciding whether the next AI bet should understand physics, not just patterns.
Related field notes
The Small Business AI Opportunity Is Hiding in Plain Sight
Small businesses are not short on ideas. They are short on capacity. AI works when it acts as a support layer around the owner — not as a replacement for their judgment.
May 22, 2026Why industrial AI needs its own org chart
IT is built for support. Operations is built for survival. Neither is structured to make AI work in the field — which is why industrial AI needs its own accountable operating model.
May 14, 2026The Next AI Shift Is Delegation
The hard question is not what AI can do. It’s what we should let it do, and how.
Working through this decision?
A short note on the AI bet you are weighing is enough to start. I usually reply within one business day.