Explainable AI for Semiconductor Manufacturing Quality

Scrapping a single wafer at the wrong moment doesn’t just hurt yield—it can cascade into missed delivery windows, rework backlogs, and expensive tool time that never comes back. Semiconductor manufacturing lives on tight tolerances, long process chains, and a hard truth: you rarely get a second chance to measure what matters once a layer is deposited.

That’s why the work highlighted in an interview with PhD researcher Amina Mević caught my attention. She’s building an explainable, multi-output virtual metrology system to predict physical properties of metal layers during semiconductor manufacturing—research done in collaboration with Infineon Technologies Austria under Europe’s microelectronics initiative. If you’re following our AI in Robotics & Automation series, this is a concrete example of what “intelligent automation” actually looks like on a fab floor: fewer blind spots, better decisions, and models that engineers will trust enough to act on.

Virtual metrology is about speed—and avoiding measurement bottlenecks

Virtual metrology uses machine learning to estimate process outcomes when direct measurement is slow, expensive, or destructive. Instead of waiting for offline metrology results (or relying on sparse sampling), you infer key quality variables from sensor and tool data in near real time.

In semiconductor production, metrology is a classic throughput constraint. Some measurements:

• Require dedicated tools with limited capacity
• Happen only after several process steps (so defects are discovered late)
• Involve destructive testing or sample-based inspection

Virtual metrology flips the timeline. You can predict film thickness, sheet resistance, uniformity, or other layer properties right after deposition—or even during the run—based on recipe parameters, chamber conditions, and sensor traces.

This matters for robotics and automation because modern fabs are already highly automated mechanically. The next bottleneck is decision automation: when to hold a lot, when to reroute, when to tweak a recipe, when to trigger maintenance. Virtual metrology gives those automated systems a faster “sense of reality.”

Why multi-output prediction is the next step up

Most production teams start with single-target models: one model predicts one metric (say thickness). But in real processes, quality variables move together.

Multi-output virtual metrology predicts several correlated properties at once, which has two big advantages:

1. Consistency: Predictions respect the fact that outputs are physically linked.
2. Operational value: Engineers don’t make decisions on one variable; they judge tradeoffs across multiple.

Mević’s focus on multi-output prediction is pragmatic. It reflects how engineers actually run a process: they’re balancing multiple specs, not optimizing a single number.

Explainable AI isn’t optional in high-stakes manufacturing

If a model can’t explain itself, it won’t get used for critical decisions. That’s not academic purism; it’s how fabs protect yield, safety, and compliance.

In the interview, Mević describes building an explainable approach that identifies the most relevant input features for multi-output predictions. That’s exactly where explainability becomes operational:

• It helps engineers validate that the model learned process truth rather than noise.
• It helps troubleshoot when predictions drift.
• It supports decisions that affect scrap risk and customer commitments.

Here’s the stance I’ll take: black-box accuracy is overrated if it slows adoption. A slightly less accurate model that engineers trust and act on can outperform a mysterious model that sits in a dashboard no one uses.

What “good explanations” look like on a fab floor

Not all explainability is useful. A bar chart of “feature importance” can be misleading if it’s unstable, correlated, or not aligned to the physics.

Explanations are valuable when they:

• Match process intuition (or clearly show why intuition is wrong)
• Stay stable across reasonable operating ranges
• Differentiate controllable vs. uncontrollable factors
• Support action (“change X”, “inspect Y”, “schedule cleaning”)

Mević’s emphasis on collaboration with industry experts is a tell that she’s optimizing for this kind of practical explainability—not just academic interpretability.

From PVD data to decisions: where physics-informed ML helps

The strongest manufacturing ML systems respect the physics, even when they’re data-driven. In the interview, Mević points to a synergy between physics, mathematics, and machine learning—working with physical vapor deposition (PVD) data and using geometric and algebraic ideas (projection operators) to improve performance and interpretability.

You don’t need to be a quantum mechanics expert to appreciate the principle: good feature selection and model structure can encode the constraints engineers already know exist.

Projection-based selection: a practical way to reduce noise

Manufacturing datasets are messy:

• Sensors drift
• Maintenance events change baselines
• Recipes evolve
• Lots differ by incoming material variation

Feature selection isn’t just “reduce dimensionality.” It’s risk management. If your model is built on fragile signals, it will fail in exactly the moments you need it.

Mević mentions applying a projection-based selection algorithm (ProjSe) aligned with process physics and domain knowledge. The deeper point for automation teams is this:

The best industrial AI models aren’t the ones with the most features—they’re the ones with the most defensible features.

In robotics and automation deployments, defensibility translates to faster sign-off, easier troubleshooting, and fewer late-night “why did the model do that?” escalations.

Time series is where virtual metrology becomes truly real-time

The next frontier is multivariate time series modeling, which Mević plans to focus on. That’s the direction most fabs are heading because tool data isn’t static—it’s a stream.

If you’re only using end-of-run summary statistics, you miss:

• Transient instabilities during ramp-up
• Chamber conditioning effects
• Short excursions that still cause defects
• Early indicators of a tool drifting out of spec

Time series models can catch these dynamics and enable earlier intervention. Done right, you can shift from “predict after the run” to predict during the run.

How time series predictions tie into automation and robotics

In the AI in Robotics & Automation context, multivariate time series virtual metrology connects directly to closed-loop control and automated response:

1. Sense: collect high-frequency signals (pressure, power, gas flow, temperature, vibration)
2. Predict: estimate multiple quality metrics continuously
3. Decide: determine whether to continue, adjust, or stop
4. Act: recipe tuning, lot holds, automatic rework routing, or maintenance triggers

This is where AI stops being “analytics” and starts being automation infrastructure.

Responsible AI and the EU AI Act: plan for compliance early

Mević explicitly mentions aligning responsible AI work with the EU AI Act principles. That’s timely for December 2025 because many industrial AI teams are now being asked the same uncomfortable questions:

• Who is accountable when a model drives a production decision?
• Can we explain the decision to auditors and customers?
• How do we monitor drift and performance degradation?
• What’s the human override path?

My opinion: compliance is easier when you design for it from day one. Retrofitting governance onto a model already embedded in production automation is painful and expensive.

A practical “responsible AI” checklist for virtual metrology

If you’re building or buying AI for semiconductor manufacturing, you want concrete guardrails. Here’s a checklist I’ve found useful:

1. Traceability: Every prediction should be traceable to model version, training data window, and tool state.
2. Explainability by design: Provide explanations that are stable and process-relevant, not just mathematically convenient.
3. Human-in-the-loop thresholds: Define when automation can act automatically vs. when it must request approval.
4. Drift monitoring: Track both input drift (sensor distributions) and output drift (prediction errors when metrology arrives).
5. Fail-safe behavior: Decide what happens when the model is uncertain or data is missing (fallback rules matter).
6. Bias and representativeness: Ensure training data covers maintenance cycles, recipe variants, and known edge cases.

These aren’t “nice to have.” They’re the difference between a pilot that impresses and a system that survives production reality.

What this interview teaches automation leaders (not just researchers)

This isn’t just a story about one PhD project. It’s a template for how industrial AI succeeds.

1) Collaboration beats brilliance

Mević highlights working closely with engineers to understand intricacies and validate models. That’s the playbook.

Industrial ML fails when it’s built in isolation. The fastest path to ROI is pairing data scientists with process engineers early, with shared ownership of KPIs.

2) Explanations drive adoption—and adoption drives ROI

Accuracy is only half the battle. If engineers don’t trust the model, it won’t influence the process, so it won’t change outcomes.

A strong virtual metrology program treats explainability as a product feature, not a research add-on.

3) Multi-output modeling matches how production decisions are made

Teams don’t manage one spec at a time; they manage a system of constraints. Multi-output predictions are aligned with the operational decision layer—especially in automated manufacturing workflows.

4) Time series is where you get from “prediction” to “control”

If your goal is intelligent automation, streaming models are the bridge to closed-loop interventions and proactive maintenance.

If you’re implementing virtual metrology, start here

If this interview has you thinking, “We should be doing this,” good. Here’s a practical starting plan that doesn’t require a moonshot budget.

A 30–60–90 day rollout plan

• Days 1–30: Define the target and data contracts

  • Pick 1–3 metrology targets with real operational value
  • Map which sensors/recipe parameters exist and how reliably they’re logged
  • Define what “good” means (error tolerance tied to spec limits)

• Days 31–60: Build a baseline + explainability layer

  • Train a baseline model and compare against simple heuristics
  • Add feature relevance explanations and validate them with engineers
  • Establish a backtesting loop using historical lots

• Days 61–90: Pilot in a decision workflow (not just a dashboard)

  • Use predictions to drive one controlled action (e.g., lot hold suggestions)
  • Set human approval gates
  • Track impact: avoided scrap, reduced cycle time, fewer metrology queues

If you can’t tie the pilot to a decision, it’s not automation—it’s reporting.

Where explainable virtual metrology is headed next

The direction is clear: faster predictions, tighter integration with automation systems, and stronger governance. Research like Mević’s is pushing on all three at once: multi-output prediction, explanation methods that engineers can use, and responsible AI alignment.

For our AI in Robotics & Automation series, I like this example because it’s honest about what makes industrial AI hard: the physics is real, the stakes are high, and trust is earned.

If you’re exploring AI-driven quality control in semiconductor manufacturing—or you’re trying to connect predictive models to robotics, dispatching, or closed-loop process control—this is a good moment to audit your own approach: Is your model fast enough, explainable enough, and governed well enough to run production?