Resist Bake Effects on Overlay in Maskless Lithography

How bakes move patterns

Every bake is a thermo-mechanical event. The substrate expands with temperature; the resist film and any underlayers expand at different rates and change stress as solvents leave or the resist chemically reacts. The result is bending and in-plane distortion that is rarely symmetric, because heating is rarely symmetric: edge heating, backside contact quality and lift-pin transfer all create gradients. When the substrate returns to room temperature on the exposure chuck, its pattern grid has shifted—usually by a combination of uniform scale change, bow-induced apparent distortion and local warp.

For a single exposure layer, none of this matters. Overlay is where it bites: layer two must land on layer one, and everything that moved the substrate between the two exposures enters the overlay error. In maskless lithography for photonics, MEMS and multi-level devices, overlay budgets of 100 to 300 nanometres are common, while an uncontrolled PEB on a thin glass substrate can shift patterns by more than a micrometre. The tool's encoder-based stage metrology is accurate; the substrate simply is not where the stage's map says it should be.

The PEB problem and the chill plate

Post-exposure bake dominates because it is the hottest, longest step and, for chemically amplified resists, functionally critical. Distortion during PEB comes from through-thickness and in-plane temperature gradients: the substrate edge heats faster than the centre when backside contact is poor, and lift-pin transfer adds a transient shock. Two wafers baked with identical setpoints but different transfer times can show measurably different distortion, which is why recipe control must extend beyond the plate to the handling.

The chill step is not neutral either. Rapid, nonuniform cooling freezes in the gradient the substrate had at the end of PEB, and edge-first cooling can bow the wafer in the opposite sense. A chill plate with worn vacuum grooves or contamination under one corner produces a distortion signature that repeats per plate but varies between plates—one of the more confusing overlay failure modes, since 'the recipe did not change' while the hardware did.

• PEB contributes the largest bake-induced overlay term in most flows
• Transfer time and lift-pin behavior are recipe parameters, not conveniences
• Chill plate condition and backside contact uniformity belong in preventive maintenance
• Distortion is repeatable per setup, which makes it measurable and partly compensable

Measuring bake distortion honestly

The distortion is measurable with modest effort. Read alignment marks or dedicated overlay patterns on a representative substrate immediately after exposure, then again after the complete bake chain, using the same setup and the same metrology. The displacement field between the two readings is the bake contribution: decompose it into translation, scale, rotation and residuals. A dominated single term points at a specific cause—pure scale change suggests uniform temperature error, edge-heavy residuals suggest plate nonuniformity or transfer asymmetry.

Once quantified, the term earns its place in the overlay error budget alongside stage accuracy, alignment error and pattern-placement contributions. Two practical cautions: measure on the production substrate stack, because distortion scales with substrate stiffness and film stress, not on a convenient dummy; and re-measure after any recipe, plate or resist change, because the term does not transfer between them.

Designing the bake chain for overlay

Where the process allows, the strongest control is architectural: bake on the same chuck geometry used for exposure, or perform exposure and development between bakes without re-clamping, so distortion never has a chance to accumulate between pattern layers. Where separate plates are unavoidable, choose proximity-style uniform heating, verify plate maps periodically, fix handling, and keep a log linking every overlay campaign to its exact bake recipe and hardware.

With those disciplines, bake distortion becomes a stable, quantified, partly compensable term rather than a random one. Some of it can even be pre-compensated: if the distortion field is repeatable, the maskless data path can apply a distortion correction to later layers, provided the compensation is requalified whenever the bake chain changes. The guiding principle is the same as anywhere in precision work—an error you can measure and hold constant is an error you can manage.