Designing a Training Cut Set

Training recovers the six milling coefficients by fitting simulated force against measured force over a set of passes. Whether a given set of cutters and passes can determine those coefficients at all is a property of the design rather than of the data quality: two specific degeneracies leave a coefficient unrecoverable however clean the samples are. Each has its own fix, and neither fix substitutes for the other.

The Two Degeneracies

What goes wrong	How it shows up in the result	What removes it
The bending-moment plane null. In rotation-averaged Mx/My one particular combination of the edge and normal shear coefficients is exactly unobservable	Shear coefficients come back orders of magnitude too large, and their signs flip between runs on the same data	Helix diversity — at least two clearly different helix angles among the cutters trained together
Slope–intercept collinearity between the shear and ploughing coefficients, which enter the fit as the slope and the intercept of the same line	The correlation still looks good. Shear comes back low by roughly a tenth, ploughing high by around a half	A spread of feed per tooth across the passes — a range of chip loads, not a range of depths

A set that fixes one and not the other still fails, and it fails in the manner of the one left unfixed. Both are properties of the cut set as a whole, so both are decided before any metal is cut.

Why the Helix Angle Is the Lever

The unobservable direction is a combination of the edge and normal shear coefficients weighted by the cosine and sine of the helix angle. Changing the helix angle does not remove that direction — it rotates it. Two helix angles far enough apart therefore give two different null directions whose only common point is zero, so the pooled fit has no null at all, and the edge coefficient becomes identifiable from the measurement alone: no torque channel, no prior value, no externally supplied phase.

A single helix angle of zero is the worst case, and misleadingly so, because it aligns the null exactly with the edge-coefficient axis. A cut set built that way returns an edge coefficient that is null-space fill rather than a value the data demands, while the normal coefficient beside it comes out as the most accurate number on the page.

Important

At a single helix angle the individual edge and normal shear values are not meaningful on their own, even when the correlation coefficient is high. A training run is judged by the force it reproduces over time — see Cutting Force and Torque — not by the magnitude of any one coefficient.

Flute Count Is Not a Lever

Adding more passes with the same symmetric multi-flute cutter buys nothing, at any feed. That sample subspace is degenerate, and more of it stays degenerate. Evenly spaced flutes also cancel the rotation-averaged transverse moment a run derives its own cutter phase from, so a set made only of symmetric passes carries no phase reference of its own. Which way it then fails depends on the channels measured: with the axial torque channel present the fit still converges and reports a high correlation over shear coefficients orders of magnitude too large; without it the fit collapses rather than degrading. Neither outcome is usable, and adding passes repairs neither. What buys identifiability is spread — in helix angle, and in feed per tooth.

Supplying the phase from outside removes that half of the problem and nothing else. EnableCwePhasePairing (API) reads each step's phase from the engagement blocks instead of bootstrapping it from the samples, for one-flute and symmetric two-flute cutters in light radial side cuts; outside those preconditions the training falls back to the bootstrapped path with a warning, so a four-flute set is not rescued by it. The null a single helix angle leaves is untouched either way.

The shipped smart-holder cut set has exactly this shape: its milling passes are one symmetric four-flute end mill at two feeds, so the set is spread in feed per tooth alone and does not determine its own edge and normal shear values — see Smart Holder Training.

Symmetric off-the-shelf cutters are nonetheless sufficient. Custom single-flute grinding is not required for either fix; a pair of stock two-flute cutters ground at different helix angles carries the same identifiability, and is what a shop can actually obtain.

Engagement: Side Cuts Rather Than Slots

The useful excitation comes from passes whose radial engagement is small enough that at most one helix flute is in contact at a time — the engagement arc narrower than the angular pitch between flutes. A full slot holds several flutes in the cut at once and averages away the very variation the fit needs. Stepping a shallow side cut along the edge of the stock produces a set where no pass is ever a slot.

Adding depth of cut is not a substitute. Depth moves the load without changing the chip load, so it does nothing for the slope–intercept collinearity that the feed-per-tooth spread exists to break.