Semiconductor fabs: tool drift, metrology gaps, cleanroom power

Cleanroom air changes run at maximum day and night

What happens. The fans of a cleanroom blow the same large amount of air through the filters whether the room is full of people and tools or empty at night. The setting was chosen for the worst case, and nobody dares to touch it, because particle counts and pressure between rooms must hold.

What it costs. Fan power is a large part of the energy of a fab, and the cooling or heating of the extra air comes on top. Filters also age faster with air that nobody needed.

Why the usual fix fails. Turning down the fans by hand at night works until a particle alarm or a pressure difference wakes the quality team. A fixed night mode ignores what actually happens in the room. A new, more efficient fan helps once and does not change the setting.

How a fixed-price remote diagnosis checks it. From logs of particle counters, pressure differences, fan speeds and occupancy or tool states I check how far the air change can follow the real need without leaving the limits. The fixed-price remote diagnosis answers whether a demand-led control with a model of particles and pressure is realistic for your rooms, and what has to be proven first.

Background: Economic predictive control

Film thickness and etch depth are measured on only a few wafers

What happens. Metrology tools measure a sample of the wafers, sometimes hours after the step. The other wafers pass the process without any measured value. When a measurement comes back out of range, many wafers have been processed with the same problem.

What it costs. Scrap wafers or rework, extra metrology capacity, and a long delay between a drift and the moment someone sees it.

Why the usual fix fails. More measurements cost tool time and cycle time. A fixed sampling plan measures the wrong wafers as often as the right ones. Reading the chamber sensors is not enough without a model that links them to the result.

How a fixed-price remote diagnosis checks it. From chamber sensor data, recipes and the measurements you already have I check how well the result of each wafer can be computed. The fixed-price remote diagnosis answers whether a computed film thickness or etch depth can fill the blind spots, and which measurements it still needs for checking.

Background: Computing unmeasured states for control

Tools drift between maintenance and the recipe stays the same

What happens. After a cleaning or a part change the tool runs well. Over the next weeks the chamber wall, the parts and the gas lines change, and the same recipe gives a slightly different result. The correction comes when metrology shows a deviation, or at the next maintenance.

What it costs. Yield that moves slowly downwards between maintenances, more scrap at the end of a cycle and more maintenance than the tool would need if the drift were known.

Why the usual fix fails. Run-to-run correction with a simple rule follows the last result but not the cause, and the measurement delay makes it slow. Shorter maintenance intervals cost production time.

How a fixed-price remote diagnosis checks it. From the history of recipes, sensor data and measured results I estimate how the tool drifts and how a recipe could be adapted from run to run. The fixed-price remote diagnosis answers whether a drift estimate is stable enough to adapt recipes within the limits of your process.

Background: Control that adapts to drift and wear

Wafer temperature is not measured directly

What happens. In thermal steps the wafer sits in a chamber with heaters and lamps. The thermocouple sits near it, not on it, and the wafer temperature depends on the position, the coating and the state of the chamber. The recipe is written for the measured value, not for the wafer.

What it costs. Uniformity across the wafer varies between runs and tools. Extra heating time makes sure of the minimum, and each step is longer than it needs to be.

Why the usual fix fails. A calibration with a test wafer holds for one state of the chamber. More sensors around the wafer add cost and still do not see the wafer. Larger tolerance in the recipe shifts the problem to the next step.

How a fixed-price remote diagnosis checks it. From the heater signals, the sensor values and the results of test wafers I check how well the wafer temperature can be computed during the run. The fixed-price remote diagnosis answers whether a computed wafer temperature can steer the heaters and shorten the step.

Background: Computing unmeasured states for control

The same recipe gives different results on different tools

What happens. A recipe was developed on one tool and is copied to others of the same type. The results differ slightly. Process engineers adjust each chamber by hand until it matches, and the adjustment is lost at the next maintenance.

What it costs. A yield spread between tools, engineers who spend weeks on matching, and production that must be sent preferably to the best chamber.

Why the usual fix fails. Offset tables per chamber pile up and nobody remembers why they exist. A tighter hardware tolerance is expensive and does not remove the ageing. Matching by hand is a job for the few people who know the tool.

How a fixed-price remote diagnosis checks it. From the sensor data and measured results of several tools I check which differences a model can explain and which parameters would have to be changed to match them. The fixed-price remote diagnosis answers whether model-based matching can replace the offset tables, and how much data it needs.

Background: Replacing hand-tuned parameter tables by a model

The ultrapure water plant is run conservatively

What happens. Ultrapure water is made with several treatment steps in a row, and the plant is set so that the water quality never drops below the limit. Pumps, membranes and regeneration cycles run on fixed settings, whatever the real demand from the fab is.

What it costs. Energy for pumping and treatment, chemicals for regeneration and water that is thrown away because the plant could not follow the demand.

Why the usual fix fails. A bigger buffer tank gives calm, but also needs space and keeps the water standing. Changing a setting by hand is risky when the water quality is an alarm criterion. Savings from a new pump are small against the settings.

How a fixed-price remote diagnosis checks it. From operating logs (flows, pressures, conductivity, regeneration times, demand of the fab) I check how far the plant can follow the demand with less energy and chemicals. The fixed-price remote diagnosis answers whether an optimised operation with a model can run within your water quality limits.

Background: Economic predictive control

Exhaust abatement runs at fixed settings whatever the load

What happens. Process gases from the tools are burned or washed before they leave. The abatement units run on fixed settings for fuel, air and water, set for the worst gas load. Most of the time the load is much lower.

What it costs. Fuel gas, water and power for treatment that the real load does not need, and wear of the burners that run at high temperature around the clock.

Why the usual fix fails. Lower settings by hand risk a failure of the abatement and a report to the authorities. A simple fixed turn-down by schedule does not know which tool is running which recipe.

How a fixed-price remote diagnosis checks it. From logs of the abatement units and the tool schedule I check how much the load changes and how the settings could follow it within the permitted limits. The fixed-price remote diagnosis answers whether a load-led control is feasible and what safety proof it needs.

Excursions in the process are found late

What happens. A tool starts to behave differently: a pressure curve looks odd, a flow is a little off. The engineers see it when an alarm limit is crossed or when metrology shows the damage. Between the first sign and the alarm, many lots have been processed.

What it costs. Scrap wafers, lots on hold and engineers who search through a pile of sensor traces afterwards to find what changed.

Why the usual fix fails. Tighter alarm limits create false alarms that people learn to ignore. Wide limits miss the slow changes. A person cannot compare hundreds of traces for every lot.

How a fixed-price remote diagnosis checks it. From sensor traces of good lots and of known excursions I check which deviations a model can flag early and how often it would raise a false alarm. The fixed-price remote diagnosis answers whether model-based detection earns its place next to your existing limits, and how it would be handed over.

Specialty gases are dosed with a safety margin

What happens. Process gases are expensive, and the dosing is set with a margin so that the process never runs short. Pressure and flow in the supply lines change with the demand of all tools, so each tool gets slightly more than it needs.

What it costs. Gas cost on every wafer, gas that is purged or abated unused, and more load on the supply and the exhaust treatment.

Why the usual fix fails. Reducing the flow by hand risks a failed process. A better flow controller improves the accuracy at the valve but not the choice of the setpoint. Tighter specifications from the process side rarely arrive.

How a fixed-price remote diagnosis checks it. From supply logs, flow setpoints and process results I check how far the dosing could be cut by a more precise control and a computed flow at the tool. The fixed-price remote diagnosis answers whether the margin can be reduced without touching the result of the process.

Tool scheduling follows dispatch rules, not an optimised plan

What happens. Lots are sent to the next free tool by simple rules: first in, first out, or by priority. In a fab with hundreds of steps and shared tools, queues build up in front of the bottleneck while other tools wait for material.

What it costs. Long cycle times, high work in progress in the fab, late lots and a lot of time spent by planners who correct the plan by hand each shift.

Why the usual fix fails. Adding a new rule solves one case and creates another. A bigger planning team does not change the principle. A forecast of the queue helps, but it does not decide which lot goes where.

How a fixed-price remote diagnosis checks it. From a lot history (step times, queues, tool states, maintenance windows) I check where the bottleneck is and how much an optimised plan could win against your rules. The fixed-price remote diagnosis answers whether a scheduling optimisation is feasible within the time you have for a decision.

Background: Optimal paths and setpoint plans

Process cooling does not hold the temperature steady at the tools

What happens. Chillers and cooling loops supply the tools with water of a set temperature. When many tools start or stop at once, the supply temperature swings. The tool then reacts and the process engineers see a drift they cannot trace to a cause.

What it costs. Yield loss from temperature variation at the tool, and chillers that run with extra reserve so that the swing stays small.

Why the usual fix fails. A larger buffer smooths the swing but needs space and money. Tighter standard controllers on each chiller do not know what the other tools will do in a minute. More reserve capacity costs energy all the time.

How a fixed-price remote diagnosis checks it. From logs of supply temperature, flows, chiller states and the tool schedule I check how large the swings are and how far they follow from the load changes. The fixed-price remote diagnosis answers whether a predictive cooling control that sees the load change coming can keep the temperature steadier.

Utility systems of a new fab are untuned at the start

What happens. When a new fab or a new hall starts, the utilities (cooling, air, water, gases) are built and tested in parts. Together, at the load of production, they run with default settings. The first months show oscillations, alarms and an energy use higher than planned.

What it costs. Delays in the ramp-up, engineers from many trades on site for months and a high energy cost until the settings are right.

Why the usual fix fails. Commissioning checklists confirm that every part works, not that the whole runs well. More engineers on site shorten the first weeks and do not leave the knowledge behind. Settings from the supplier are made for the average site.

How a fixed-price remote diagnosis checks it. From the design documents, controller settings and first operating logs I check which loops cause the oscillations and what a systematic tuning of the plant would involve. The fixed-price remote diagnosis answers where a model-based pre-tuning saves time in the ramp-up.