AuxoAI

⚠

The Challenge

💡

The Solution

Increase in total cycle time and disrupted production flow due to frequent wafer holds

›››

Automated real-time prediction of wafer disposition, covering rework, re-measure, or continue to the next step

Manual disposition required engineers to analyse historical and tool data for each hold, creating delays and inefficiencies

›››

A machine learning model trained on historical hold and tool data from metrology, inspection and process tools

Inconsistent decisions stemming from subjective, expertise-dependent evaluations impacted yield and rework accuracy

›››

Multiple tool and process data sources integrated into a single view to support model-driven decisioning

Wafer movement delays slowed downstream processes and affected overall throughput

›››

Manual classification replaced with AI-assisted dispositioning to reduce delays and engineering effort

The Full Context

Holds are how a fab protects yield. A wafer trips a measurement limit, it stops, and someone decides what happens next. The cost sits in that decision, not in the hold itself. An engineer had to pull historical hold records, cross-reference tool data, and form a judgment for every lot that stopped. Wafers waited while that happened.

Because the judgment depended on who was reviewing it, similar anomalies did not always get similar answers. That variability showed up downstream as unnecessary rework, avoidable scrap risk, and yield that was harder to explain after the fact.

AuxoAI worked with the manufacturing team to build Wafer Ops AI, a prediction layer that reads the same signals an engineer would and returns a recommended disposition as soon as the hold is raised. The engineer stays in the loop and keeps the decision. What changes is that the analysis is already done.

The Approach

AuxoAI sequenced three phases so that each one had to earn engineering trust before the next began.

Unified the Data the Decision Needs

Metrology, inspection and process tool sources were integrated alongside historical hold outcomes, so every prediction draws on the same evidence an engineer would gather manually.

Trained the Disposition Model

A machine learning model learned from historical hold and tool data to predict the right disposition for a wafer: rework, re-measure, or continue to the next step.

Put It Where the Work Happens

Predictions surfaced in real time at the point of hold. Manual classification gave way to AI-assisted dispositioning, with engineers reviewing and confirming rather than starting from raw data.

The Impact

Wafer Ops AI delivered a 3% reduction in total cycle time and returned 2.5 hours per day of engineering productivity, with increased throughput and efficiency across the line.

The hours came back first. Engineers stopped rebuilding the same analysis for every hold and moved to reviewing a recommendation, which freed time for the process work that only they can do.

Cycle time followed. Wafers spent less time waiting for a decision, and downstream steps stopped absorbing the delay from upstream holds.

The quieter gain is consistency. Dispositioning now starts from the same evidence and the same logic every time, which makes rework accuracy easier to hold steady and yield outcomes easier to trace back to a cause. The same pattern applies to any decision point in the fab where a trained judgment is being repeated against tool data.

Project Highlights

Solution

Real-time wafer disposition prediction

ML model on historical hold and tool data

Integrated metrology, inspection and process sources

AI-assisted dispositioning in the engineer workflow

Results

✓3% cycle time reduction

✓2.5 hrs/day productivity gain

✓Increased throughput and efficiency

✓Consistent, evidence-based hold decisions