In Engine Prognostics, the Model Isn't the Hard Part. Knowing What Matters Is.

A personal take on why engine prognostics needs a continuous loop, not just more data or a bigger model.

For me, the interesting part of applying AI to aircraft engines was never the model. On the V2500 and CFM56, decades of shop visits already told us which parameters correlate with deterioration and cost. On the PW GTF and CFM LEAP, that answer doesn't exist yet, and it can't be borrowed from the mature fleet. The work that matters is building a loop that discovers what actually drives the outcome, and keeps re-learning as real flight hours come in.

When people talk about AI in aircraft engines, they usually frame it as a modeling exercise. Collect more data, train a bigger model, watch the prediction improve. That framing holds up when we already understand the relationship between inputs and outcomes. It does not hold up when we don't, and for a meaningful slice of today's fleet, we don't.

On the V2500 and CFM56, the industry has decades of operational and shop-visit history behind it. That history did something specific: it told us, parameter by parameter, what correlates with component deterioration, removals, repairs, scrap rates, and cost. By the time I started looking at this problem, the modeling question on a mature engine was mostly already answered, because the harder question, which variables matter, had been answered first, one shop visit at a time, over many years.

The PW GTF and CFM LEAP don't get that shortcut. And here is where I want to be specific about something people get wrong: the goal was never to take a V2500 or CFM56 model and point it at a GTF or LEAP. Engine architecture, thermal management, and failure modes differ enough that a parameter weighting learned on one family does not transfer cleanly to the next. What I needed to figure out, from the engineering and operational data that does exist, was which parameters matter most for the new engine, and how their relative importance shifts as more of the fleet accumulates flight hours.

That question got sharper this year. CFM secured FAA and EASA certification for the LEAP-1B's high-pressure turbine durability kit in July 2026, aimed at roughly doubling time on wing in hot and harsh environments, with fleet cutover expected in 2027. Pratt & Whitney's GTF Advantage is now certified on the A320neo family and entering service through 2026. Both will shift the reliability profile of engines already flying, which means any parameter map I build today has to keep updating as those changes reach scale, not sit frozen the day it ships.

That is where the real AI loop starts

The candidate variables are not a mystery. Flight hours and cycles (FH/FC), EGT margin, thrust settings, environmental conditions, vibration, oil consumption, component history, maintenance events: none of this is secret information. What is not known in advance, for a new engine type, is which combination of these explains what happens to the engine, and how that combination changes as the fleet matures.

That is the reframe I keep coming back to. It is not a question of whether you have the data. It is a question of which data explains the outcome, and that is an empirical question. You only answer it by testing predictions against reality.

So the loop I trust looks like this:

1. Make a prediction.
2. Operate the engine.
3. Observe the outcome.
4. Compare prediction against reality.
5. Re-evaluate the parameters and their weightings.
6. Improve the model.
7. Run it again.

Run that loop long enough and something happens that a one-time training run cannot replicate. The system stops just accumulating data and starts accumulating understanding: which parameters move the outcome, and by how much, under which conditions. That is the part a static model, no matter how much historical data it started with, cannot give you on its own.

Why this matters now

I think this matters more this year than it did two years ago, precisely because the GTF and LEAP fleets are old enough to be interesting and young enough to still surprise us. The durability kit and the GTF Advantage are not small updates. They change failure modes in engines that are already flying, on a timeline that does not wait for the next multi-year data cycle to catch up.

If your model was trained once and shipped, it is already behind. If your loop is running, it absorbs the new reliability data as it arrives and re-weights accordingly. That difference is not academic. It shows up as the gap between a shop visit forecast that is still useful in year three and one that quietly stopped being accurate somewhere around year one.

Where I think the real advantage sits

AI models will keep getting cheaper and easier to access. That is a leveling force, not a competitive edge. What I think compounds is the ability to continuously figure out what matters, check it against real engine behavior, and feed that back into the prediction before the next shop visit.

For aviation, I do not think the long-term moat is the model. I think it is the accumulated intelligence about which variables move the outcome, built one shop visit at a time. That kind of intelligence compounds with every flight hour and every shop visit in a way a licensed model never will, because it is earned, not purchased.

If that is right, the infrastructure question changes too. It stops being "which model do we license" and becomes "do we have a closed loop that captures every prediction, every outcome, and every re-weighting, so the system gets smarter with each cycle instead of staying frozen at whatever it knew on day one." That is the same discipline behind Aviation Record Intelligence: back-to-birth records, EGT margin trends, and component history are not one-time inputs. They are the ongoing feedback that keeps the parameter map honest as engines like the GTF and LEAP rack up real operating history.