Explainable Causal-Neurosymbolic AI for Process Analysis and Optimization in Semiconductor Manufacturing

Semiconductor manufacturing processes are among the most complex engineered systems, involving intricate interactions between materials, equipment, process recipes, environmental conditions, operational procedures, and human decisions. Modern Artificial Intelligence (AI) methods increasingly support process monitoring, yield prediction, virtual metrology, and defect analysis by learning patterns from large volumes of sensor and manufacturing data. However, most existing approaches remain predominantly statistical in nature, identifying correlations rather than understanding the underlying mechanisms that govern process behaviour. As a result, they often struggle to explain process deviations, support root-cause analysis, predict the effects of interventions, or generalize robustly when process conditions change.

At the same time, semiconductor organizations possess rich sources of knowledge beyond raw data. Process engineers capture expertise in manuals, process flows, equipment documentation, best practices, physical models, and tacit operational knowledge. In parallel, manufacturing data itself contains hidden process signatures, recurring behavioural patterns, and latent relationships that are not explicitly documented by experts. Current AI solutions typically exploit only one of these knowledge sources, resulting in fragmented models that either lack domain understanding or fail to leverage the full informational content of the data.

This PhD aims to establish a novel Causal Inference and Neuro-symbolic AI  with the aim to extract expert knowledge, machine-discovered process knowledge, and causal reasoning embedded into a common graph-based representation for semiconductor manufacturing.

The first pillar focuses on constructing rich semantic representations of semiconductor processes. Ontologies, Knowledge Graphs, and semantic models will be used to formally capture process entities, process flows, equipment relationships, parameter dependencies, engineering constraints, and expert knowledge. Large Language Models (LLMs) will be investigated as enablers to accelerate ontology creation and maintenance by extracting structured knowledge from technical documents, reports, process descriptions, and engineering documentation.

The second pillar focuses on discovering and representing knowledge that emerges directly from manufacturing data. Rather than treating sensor streams, metrology measurements, and production histories purely as numerical inputs, the research investigates how meaningful process signatures can be extracted and transformed into graph representations. Techniques from representation learning, graph learning, graph walks, temporal pattern mining, and neuro-symbolic reasoning will be explored to identify recurring process behaviours, latent process states, defect fingerprints, equipment characteristics, and hidden relationships among process entities. These learned structures become complementary knowledge assets that extend and enrich expert-defined process graphs.

The third pillar introduces causal reasoning as the unifying mechanism between both knowledge worlds. The goal is not merely to predict process outcomes, but to model and reason about cause-effect relationships throughout the manufacturing flow. By combining expert-defined dependencies, learned process signatures, observational process data, and graph-based representations, the framework will investigate how causal hypotheses can be generated, refined, validated, and exploited. This enables reasoning about process behaviour at different levels of abstraction, from physical process steps to system-wide manufacturing interactions.

A key objective is the creation of explainable reasoning graphs that make both human knowledge and machine-discovered knowledge transparent to engineers. Predictions, diagnoses, and optimization recommendations should be traceable through semantic relationships, discovered process signatures, and causal pathways. This provides process engineers with understandable explanations of why certain outcomes occur, which process factors contributed most strongly, and what interventions are expected to have the largest impact.

The PhD research will be evaluated on semiconductor manufacturing applications such as virtual metrology, yield prediction, process drift detection, defect Pareto analysis, root-cause investigation, and process optimization. Rather than delivering purely predictive models, the research aims to create intelligent process reasoning systems capable of answering questions such as:

The outcome of this PhD will be a next-generation causal-neurosymbolic system that combines semantics, graph representations, representation learning, causal reasoning, and explainable AI into a trustworthy foundation for semiconductor process intelligence and optimization.