LFoundry: interpretable anomaly detection in semiconductor manufacturing
An industrial AI solution to increase quality, reduce scrap and support process experts in the early identification of anomalies, drift and production deviations.
Context
LFoundry operates in a highly complex production environment, typical of semiconductor manufacturing, where process stability, product quality, scrap reduction and the ability to intervene promptly are critical factors.
In this kind of environment, the data generated by factory systems, equipment, lots, recipes, technologies, process measurements and production planning is a fundamental resource — but an extremely complex one to exploit. The expected behaviour of a process is neither unique nor static: it changes with the technology, the recipe, the machine state, the operating conditions, maintenance interventions and the natural drift of production processes.
In this scenario, LFoundry partnered with Statwolf to develop an interpretable anomaly-detection solution, designed to fit into the factory’s existing digital ecosystem and support process experts in identifying, understanding and managing anomalies.
The challenge
LFoundry wanted to strengthen the monitoring of key processes with an AI approach able to promptly intercept drift, anomalies and unexpected behaviour, before they impacted quality, yield, scrap or operational continuity.
The challenge was not simply applying an anomaly-detection model to factory data. In an industrial environment as complex as semiconductor manufacturing, a single model is not enough: different operating conditions, recipes and technologies — plus variations tied to maintenance or configuration changes — require specific models, monitored and updatable over time.
The goal was to build a solution able to integrate existing data and systems, govern thousands of models in parallel, make anomaly scores interpretable, collect feedback from experts and keep the models reliable through MLOps practices.
Solution objectives
- Integrate with existing factory systems, including MES, production planning, process databases, lots, recipes and technologies.
- Identify anomalous behaviour against the correct baseline for each specific production context.
- Make anomaly scores interpretable, highlighting the variables and conditions contributing to each alert.
- Support process experts in root cause analysis and in prioritising interventions.
- Manage thousands of models in parallel, each tied to specific operating conditions.
- Continuously monitor model performance through MLOps practices.
- Collect user feedback to progressively improve the system, evolving from an unsupervised approach toward semi-supervised logic.
The approach
Statwolf worked with LFoundry on an end-to-end approach, combining data integration, industrial AI, interpretability, MLOps and domain knowledge. The heart of the project was bringing AI into the real complexity of the fab.
The solution was designed to interoperate with the systems already in place at LFoundry, connecting data from MES, production planning, process systems, lot information, recipes and technologies. Without a reliable connection to the production context, anomaly detection risks producing isolated signals that are hard to interpret and of little use in decision-making.
The solution is not based on a single model, but on an architecture able to run thousands of models in parallel. Each model is tied to specific operating conditions, so that observed behaviour is compared against the correct baseline for that context — allowing real anomalies to be distinguished more accurately from normal process variability.
A central element is continuous model monitoring. On the factory floor, data behaviour can change rapidly following maintenance, operational variations, recipe changes, technology changes or progressive drift: the solution includes MLOps components dedicated to performance monitoring, drift detection and model readjustment whenever the baseline is no longer representative.
Alongside this, Statwolf and LFoundry worked on collecting feedback from process experts: confirmations, corrections and validations progressively enrich the system, turning an initially unsupervised approach into a semi-supervised one — closer to production reality and more effective over time.
The solution
An interpretable, scalable anomaly-detection system integrated with the factory’s real processes, built on five main components.
Integration with factory systems
Integration with the existing digital ecosystem: MES, production planning, process data, lots, recipes, technologies and operating conditions. Every analysis is tied to the correct production context, and the data foundation builds coherent datasets for analysis.
Modelling at scale
Different technologies and recipes run in parallel on multiple machines, each with its own sub-systems. The solution manages thousands of models in parallel, each tied to specific combinations of production conditions, comparing observed behaviour with the expected baseline.
Interpretable anomaly detection
The models identify significant deviations and produce anomaly scores used to prioritise events and situations to analyse. Interpretability links each alert to the variables that contributed most and to the production context in which it occurred.
MLOps and model lifecycle
MLOps components monitor the models in production: performance, drift, sudden changes due to maintenance, operational variations or process changes. When the reference baseline is no longer representative, the system supports model readjustment.
Feedback loop and continuous improvement
Validations, confirmations and corrections from process experts feed a continuous-improvement cycle, evolving the system from an unsupervised approach toward semi-supervised logic.
Why Statwolf
The LFoundry case captures Statwolf’s positioning in industrial AI well: not just algorithm development, but domain knowledge, the ability to make AI interoperable with the systems already present in the fab, scalable across thousands of models, monitorable through MLOps and usable by process experts.
The value of the solution lies not only in detecting anomalies, but in turning AI into an operational tool — integrated into real processes and understandable for those who make technical decisions on the factory floor.
- Interoperability with MES, production planning and existing data systems.
- Anomaly detection on complex industrial data.
- Thousands of models managed in parallel.
- Interpretable anomaly scores.
- Continuous monitoring through MLOps.
- Handling of drift, maintenance and sudden changes in operating conditions.
- Feedback collection and continuous improvement.
- Support for quality, scrap reduction and production stability.
LFoundry strengthened the monitoring of its production processes with an interpretable anomaly-detection solution, integrated with existing systems and scalable across thousands of models. Statwolf enabled an AI and MLOps architecture that supports quality, scrap reduction, drift management and continuous improvement, turning complex industrial data into operational insight for process experts.
What they say about the project
“In semiconductor manufacturing, anomaly detection cannot be limited to flagging that something is out of the norm. The real value lies in making the result interpretable: understanding which variables contribute to the anomaly, separating noise from relevant signals, and turning the model into a tool that speaks the language of process experts. This was possible thanks to close collaboration between Statwolf and our internal data-science team, combining algorithmic skills, domain knowledge and an understanding of factory systems.”
“The project required strong coordination across different skill sets: data science, process knowledge, data management and factory operational goals. The value of working with Statwolf was doing it in a structured way, turning a complex R&D effort into a concrete solution, progressively validated with LFoundry’s experts.”
Reference
The activities described were partially developed within the AIMS5.0 — Artificial Intelligence in Manufacturing leading to Sustainability and Industry 5.0 project.
