Case study · precision manufacturing

The controller was causing the defect it was trying to prevent.

On a precision machining line, an adaptive feed control was slowing the tool to protect each part, and the data told a different story.

Modeled in simulation

No new hardware

The challenge

A precision manufacturer was losing high-value parts to sidewall defects: scrap and rework on a cut that has to come out right the first time. The machine already ran an adaptive feed control that slows the tool whenever cutting forces rise, on the assumption that slower is safer.

What the data revealed

It wasn’t. Reading the machine’s own sensor streams, Owl AI found the controller was dramatically over-cautious: in more than 90% of its slowdowns the force never came close to the machine’s safety limit. Worse, the prolonged slow cutting was itself eroding the sidewall. In this data, sidewall load during the crawl ran roughly 10× higher than at a normal feed. The logic meant to protect the part was quietly degrading it: a pattern invisible in any single cut, but unmistakable across millions of seconds of data.

The cutting force the controller watches stays low and flat across feed speeds, while sidewall wear spikes sharply at the crawl — The counterintuitive insight: the force the controller watches stays low across speeds, while sidewall wear spikes at the crawl. Forces shown relative.

What we built

From that historical data alone, with no new sensors and no production downtime, we built a digital twin: a closed-loop simulator of the machine, its controller, and the cut outcome. Against that twin we tested a re-tuned control policy in two forms: a simple rule any PLC can run, and an ML risk-weighted version for finer targeting. The simple rule captured nearly all of the benefit.

A closed-loop digital twin: existing machine sensors feed a twin of machine, controller and outcome, which yields a re-tuned policy that deploys back to the machine — How it works: a digital twin built from existing data. Machine → twin (machine + controller + outcome) → re-tuned policy → deploy.

The results, in simulation

Up to ~60% less sidewall wear on the modeled cuts.
Stayed within the machine’s hard force limit on every modeled cut, peaking at ≤ ~68% of it (0% breach).
On the cuts the policy acts on, ~44% less time in the deepest part of the cut, recovering capacity as a by-product (mean feed eased from ~6 to ~9).

The re-tuned policy shifts cutting time away from the crawl toward a normal feed — The fix: ease the crawl up toward a normal feed, shifting time out of the danger zone (modeled).

Histogram of peak cutting force per cut as a share of the hard limit; every cut sits well below the limit, leaving a large unused safety margin — Safety: under the eased policy, every modeled cut’s peak force stayed inside the machine’s hard limit, shown as a share of that limit. The shaded band is unused safety margin (0% breach).

Time in the critical zone, today equals 100, drops to 56 under the re-tuned policy — A by-product: less time in the critical zone on the cuts the policy touches (today = 100; modeled).

What we don’t hand-wave

These are modeled results, and we say so. The same digital twin defines a live, on-machine validation step that confirms the gains before anything ships to production. We separate what’s proven from what’s projected, every time.

Proof points (modeled in simulation): ~60% less sidewall wear · ≤68% of the hard limit, 0% breach · ~10× more wear at the crawl · 0 new sensors required.

Think your process is leaving margin on the floor?

Let’s look at your data.

Talk to us →