The mdsim datasets are synthetic, anonymized representations of data and domain knowledge found in a MEMS sensor development workflow for a fictional MEMS sensor product. They emulate the structure and heterogeneity of practical use-case data and include intentionally designed pitfalls that commonly arise when integrating information across sources. The datasets do not model physical MEMS behavior.

• Datasets/manufacturing_*.csv contains in-line manufacturing data for 2 different part types (PX1 & PX2). For each part type there are 1000 wafers simulated, each holding 25 parts
• Datasets/manufacturing_EOL_*.csv contains EOL testing data for the parts PX1 & PX2
• Datasets/engineering.csv contains in-line data for a new part type variant PX2V2. There are 10 wafers simulated, each holding 25 parts
• Datasets/engineering_EOL.csv contains EOL testing data for the part PX2V2
• Datasets/assembly_map.csv is a mapping table, tracking what parts are assembled into which sensor. There are 25000 units of sensor SY1 assembled.
• Datasets/Final_testing.csv contains the final test results for the sensor SY1 units

• engineering_coordinates.png resp. manufacturing_coordinates.png visualize wafer coordinate grids used in the simulation for engineering resp. manufacturing
• DAGs/[name]_dag.svg visualizes the directed acyclic graph (DAG) used for simulation of the respective part/sensor
• mdsim_ontology.svg graph visualization of the mdsim ontology modeling the simulated data, a simplified version of the mdev ontology

• mdsim_ontology.ttl RDF definition of the mdsim ontology
• engineering_mapping_example.ttl Example RML mapping for the engineering.cvs
• SHACL/gridalignment.ttl Example SHACL shapes for flagging wafers & parts with coordinates not matching the ontology standard
• SHACL/units.ttl Example SHACL for flagging non qudt units

The following conflicts have been injected into the datasets:

• products are uniquely identified by usn NOT by prod_name
• parts are uniquely identified by x_pos, y_pos, wafer_id, they have no ID
• unit:W is the only valid qudt unit in engineering and manufacturing. All others need preprocessing / conversion
• PXF_THK_ADJ uses µm in manufacturing & nm in engineering
• manufacturing adds a new measurement parameter Bond_ZOffset to the processing of part PX2
• manufacturing/engineering use prod_name for part types, assembly uses sub_prod_name. (prod_name in assembly describes the sensor type), FT uses prod_name for the type but has no part type info
• engineering uses a different wafer coordinate grid than manufacturing (not center)
• engineering uses Linux timestamp, manufacturing uses datetimestamp
• EOL testing records start & duration, final testing records start & end

The simulation framework implements a DAG-based causal model to capture dependencies between process parameters. Each parameter node in the DAG can serve as either an exogenous input (controllable parameter) or an endogenous output (measurable parameter derived from upstream causes). The causal structure enables systematic propagation of parameter values through the manufacturing process chain. Timestamps and other metadata are tracked and propagated as part of the simulation process as well.

Parameters are categorized into two types:

1. Controllable parameters: Design variables sampled from specified bounds using Latin Hypercube Sampling (LHS) to ensure orthogonal coverage of the design space
2. Measurable parameters: Endogenous variables computed from upstream causal relationships

The causal computation proceeds as follows:

1. Initialization: Sample controllable parameters (LHS)
2. Topological traversal: Compute measurable parameters in topological order
3. Validation: Check parameter values against specification limits
4. Postprocessing: Split and transform into data structures resembling real datasets. Introduce common SICs.

The DAG structure guarantees that all parameters can be computed in a valid topological order, where each node's value depends only on previously computed upstream nodes. This ensures deterministic, reproducible propagation without circular dependencies.

Each directed edge $(P_i \rightarrow P_j)$ in the causal graph represents a functional relationship between parent parameter $P_i$ and child parameter $P_j$. For the mdsim, we use linear models as the default.

For a linear edge relationship, the contribution from parent $P_i$ to child $P_j$ is computed as: $$c_{i \rightarrow j} = w_{ij} \cdot v_i$$ where:

• $v_i$ is the value of parent parameter $P_i$
• $w_{ij}$ is the edge weight (default: 1.0)

For a child parameter $P_j$ with multiple parents ${P_1, P_2, ..., P_k}$, the value is computed as: $$v_j = b_j + \sum_{i=1}^{k} w_{ij} \cdot v_i + \epsilon_j$$ where:

• $b_j$ is the intercept (bias) term for parameter $j$ (default: 0.0)
• $w_{ij}$ are the edge weights from each parent
• $\epsilon_j \sim \mathcal{N}(0, \sigma_j^2)$ represents measurement noise