Concept · EDA Tools & Flows
Model drift
A fall in the accuracy of a machine learning model or AI-assisted tool over time, as the data it meets in use moves away from the data it was trained and qualified on, or as the model itself is changed.
- Also known as
- Concept drift, Data drift
- Type
- Concept
- Domain
- EDA Tools & Flows, AI in EDA
Overview
Model drift is the gradual loss of accuracy of a model once it is in use. A tool can pass its qualification on one set of designs and then give worse results as new designs, process nodes, libraries or coding styles reach it, or after its supplier changes the model behind it. Nothing fails loudly; the outputs simply become wrong more often.
How it works
A model learns patterns from the data it was trained on. When the inputs it meets in use differ from that data, for example a new process node with different timing behaviour, the patterns no longer fit as well and its predictions or suggestions lose accuracy. This is often called data drift. When the relationship between inputs and the right answer itself changes, it is called concept drift.
Drift can also come from the tool side. A supplier may retrain or replace the model, change its settings or move where it is hosted, and the same input then gives a different output.
Drift is found by monitoring rather than at qualification: rerunning a fixed set of reference designs with known answers on a schedule, tracking how often engineers override or correct the outputs, and comparing results before and after any change to the tool or its inputs.
Examples
- A timing prediction model qualified on one process node that is less accurate on the next.
- A code suggestion tool whose accuracy drops after its supplier updates the model.
- A failure triage tool that ranks causes less well once a new verification methodology changes the logs it reads.
- A rising rate of engineer overrides showing that a step's outputs are less often right.
Why it matters
In a chip design flow, a drifting tool can let errors through quietly, so the independent checks and human review around it carry more of the load without anyone noticing. Teams that adopt AI-assisted steps watch for drift as part of monitoring, and treat a change of model, settings, hosting or input domain as a reason to requalify the step before relying on it again.
It matters most to: AI-for-EDA Machine Learning Engineer, Benchmark & QoR Analysis Engineer, CAD / Design Methodology Engineer, Design Analytics & Dashboard Engineer, EDA Release & Quality Engineer
Common confusions
- Not the same as Sensor drift
- Sensor drift is a slow change in a physical sensor's calibrated output. Model drift is a loss of accuracy in a software model as its inputs or the model itself change.
- Not the same as Tool qualification
- Qualification measures a tool once against reference designs. Drift is what happens after qualification, and monitoring is how it is caught.
Related standards
- ISO/IEC 42001 (artificial intelligence management system)
- ISO/IEC 23894 (guidance on AI risk management)
Sources
- Learning under Concept Drift: A Review (IEEE Transactions on Knowledge and Data Engineering) (opens in a new tab), arXiv (University of Technology Sydney and University of Porto authors), 2026-11-10
- Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions (opens in a new tab), US Food and Drug Administration, 2026-11-10
- Concept drift (opens in a new tab), Wikipedia, 2026-11-10
Related terms
Part of the EDA Tools & Flows glossary, with every term in the topic and the lifecycles, tools and standards that use them.
Where it appears on the Hub
No other page of the Hub uses this term yet.
Suggest an improvement
Suggest an improvement to this entry
A clearer definition, an example, a correction or a source an editor can check. Editors review every suggestion before the entry changes.
Last reviewed . Written by The Semiconductor Hub editorial team. The full glossary