Semiconductor tools: stages settle slowly, qualification is late

Positioning stages settle too slowly for the throughput target

What happens. The stage moves a wafer or a mask to the next position and then has to rest within a tiny tolerance before the process can start. Vibration of the frame and of the stage itself ring on after the move. The engineers wait for the vibration to die, or they tune the controller carefully until the settling time is still too long.

What it costs. Every move that waits costs throughput of the whole tool, and throughput is what the tool is sold on. Longer settling also eats the margin for the next generation with higher accelerations.

Why the usual fix fails. A stiffer mechanism is heavier and moves slower. Tuning a standard controller more aggressively starts the vibration again. Hand-made feedforward tables work for one operating point and drift with temperature and wear.

How a fixed-price remote diagnosis checks it. From measured stage data (position error, commanded motion, frequency response, accelerations) I check where the settling time goes and how much a better model, feedforward and estimation can win. The fixed-price remote diagnosis answers whether the target is within reach with control, or whether the mechanism has to change.

Background: Precision control of robots and machine tools

New tool types need a long qualification before they ship

What happens. A new tool is built, then tuned on the real hardware, step by step, with the few engineers who know the subsystems. Each subsystem has its own controller, the tool as a whole behaves differently from the parts, and every change triggers a repeat of the tests.

What it costs. Weeks of engineers and test-bench time per tool, a delayed first delivery and a ramp-up that is slow because every new unit again needs care by hand.

Why the usual fix fails. Copying the parameters of the previous tool brings a head start, but the new tool is not the old one. More test time costs the same weeks again. Simulation models exist, but often only for the design, not matched to the real tool.

How a fixed-price remote diagnosis checks it. From design models, drawings and the test logs of one existing tool I check how far a simulation can be matched to the hardware, so that controllers are pre-tuned before the first run. The fixed-price remote diagnosis answers where that saves tuning time and where the real tool must still be measured.

Control engineers for deposition equipment are hard to find

What happens. Deposition tools need people who know the process and the control engineering behind it: temperature, pressure, gas flows, plasma and the loops between them. Such people are rare. The role stays open for months, while the team builds the next tool generation with the staff it has.

What it costs. Controller tuning and model work wait in the queue, tools go out with conservative settings, and the experts who are there carry too much of the load.

Why the usual fix fails. Hiring takes long, and a new colleague needs a year to know the tool. Training a process engineer in control is slow. A consultant working in the team for months brings the question of status risk with it.

How a fixed-price remote diagnosis checks it. From your control design, loop recordings and open questions I check which model-based control task can be handed over cleanly, with a model and a description your team can maintain. The fixed-price remote diagnosis shows which part can be done remotely and what the result would look like.

Process models for new equipment are slow to build and to trust

What happens. Developers of new tools want models that combine heat, flow, plasma, mechanics and electrics, to try designs before hardware exists. The models take long to build, run slowly, and the engineers do not know how far to trust them. Learning methods are added to fill gaps, with data from few experiments.

What it costs. Prototypes that could have been avoided, design loops that take weeks, and models that sit unused because nobody validated them against the real tool.

Why the usual fix fails. A bigger simulation takes more computing time, not more insight. A purely learned model fits the data and fails outside it. Handing the task to a specialist team for each field leaves the combination unsolved.

How a fixed-price remote diagnosis checks it. From your simulation set-up, a few measurements and the question the model must answer I check which model detail is needed, how to match it to measured data and where it will not hold. The fixed-price remote diagnosis answers whether a fast, trusted model for design and control is within reach.

Large nonlinear optimisation of coupled physical effects does not converge

What happens. Developers of optical and plasma equipment optimise designs where light, heat, material and electrics act on each other. The optimisation problem is large and nonlinear. Solvers stop at poor points, need very long, or fail when one of the parameters moves a little.

What it costs. Weeks of computing and many restarts, designs that stay at a local optimum, and few people who can find out why a run failed.

Why the usual fix fails. A faster computer shortens each run, but not the number of attempts. A standard solver with default settings is built for something else. A general tool without knowledge of the structure of the problem does not find the structure either.

How a fixed-price remote diagnosis checks it. From the model equations and a typical failing run I check scaling, the structure of the problem and the choice of the method. The fixed-price remote diagnosis answers whether the problem can be formulated so that a solver converges reliably, and how much time that would save.

Background: Economic predictive control

Fast inspection mechatronics shakes and loses accuracy

What happens. Inspection systems move optics, masks or samples quickly and exactly, then have to hold still to measure. The mechanism has resonances that ring on. The more you speed up the moves, the stronger the vibration at the moment of measurement.

What it costs. Waiting times in every cycle reduce the number of parts inspected per hour, and vibration lowers the quality of the measurement itself.

Why the usual fix fails. Heavier damping parts and stiffer frames add mass and cost. A notch filter in the standard controller removes one resonance, not the changes from one unit to the next. Tuning every unit by hand takes the time of a specialist.

How a fixed-price remote diagnosis checks it. From frequency responses and recorded moves of your mechatronics I check which resonances matter, how much damping a model-based controller can add, and what remains for the mechanics. The fixed-price remote diagnosis answers whether the cycle can be faster without losing accuracy.

Background: Precision control of robots and machine tools