Instrumented Datum Definition

• Instrumented Datum is a structured data object that augments raw sensor observations with a case-specific mechanistic model, explicit uncertainty quantification, and executable counterfactuals.
• It enables targeted interventions using Pearl's do-operator and supports rigorous verification and validation processes across diverse scientific domains.
• The framework integrates parameter extraction, uncertainty propagation, and counterfactual generation to enhance causal inference and mechanistic supervision in machine learning.

An Instrumented Datum is a structured data object in scientific machine learning augmenting raw observations with a fully specified, case-dependent mechanistic model, explicit uncertainty quantification, and an executable, auditable counterfactual family. Unlike standard observational datasets, which capture only what occurred, or template synthetic datasets, which are constrained to simulator-specific templates, instrumented data encodes the explicit causal structure and permits targeted interventions on parameters via Pearl's do-operator. Instrumented data is central to mechanistically supervised learning, supports causal inference, and underpins rigorous validation and verification (V&V) workflows across domains including computational biology, climate, materials, fluid mechanics, and medical imaging (Wilke, 5 Jun 2026).

1. Formal Structure of an Instrumented Datum

Every Instrumented Datum is defined as the 6-tuple:

$D_i = (I_i, M_i, \eta_i, u_i, q_i, v_i)$

where:

• $I_i$: raw sensor observation (e.g., image).
• $M_i$: mechanistic model inferred from $I_i$, encapsulating geometry ($\Omega_i$), constitutive law ($\sigma_i$), boundary/initial conditions ($\partial \Omega_i$, $u_{0,i}$), forcings ($f_i$), and the numerical solver ($\mathcal{S}$).
• $I_i$0: explicit confounders (“outside” parameters) not contained in $I_i$1.
• $I_i$2: solver response (e.g., a field solution).
• $I_i$3: derived quantity of interest (e.g., stress, temperature).
• $I_i$4: V&V record, including verification logs, convergence and residual tests, and expert sign-off.

Additionally, two core features are attached:

• Uncertainty representation $I_i$5: push-forward of parameter uncertainties through the solver.
• Executable counterfactual family $I_i$6: each element generated by applying an interventional do-operator on any parameter in $I_i$7.

2. Extraction, Propagation, and Counterfactuals

Parameter extraction from the observation proceeds via a perception/extraction operator:

$I_i$8

For every parameter $I_i$9 (mechanistic or confounder), outputs include:

• Point estimate $M_i$0.
• Uncalibrated interval $M_i$1.
• Self-reported parameter density $M_i$2.
• Categorical confidence $M_i$3 (e.g., material class).

Uncertainty propagation employs the joint density:

$M_i$4

This gives a distribution over $M_i$5 pairs, which—via repeated extractions and calibration—can be decomposed into aleatoric ($M_i$6, irreducible) and epistemic ($M_i$7, reducible) uncertainties (the calibration protocol remains open).

Counterfactuals are generated by applying Pearl's do-operator to parameters in $M_i$8:

$M_i$9

The result is an audited, executable family $I_i$0, each reflecting specific, controlled interventions.

3. Component Explanations and Roles

A breakdown of components and their significance:

Component	Definition	Context/Significance
$I_i$1	Raw sensor measurement	Anchors simulation in real-world data
$I_i$2	Case-specific mechanistic model	Captures geometry, physics, boundary and initial conditions
$I_i$3	Explicit confounders	Parameters external to $I_i$4; extend model scope
$I_i$5	Solver response	Numerical solution enforcing model and data
$I_i$6	Downstream quantity of interest	Target for supervision, evaluation, or surrogate training
$I_i$7	Full V&V record	Auditable, enables verification and expert validation

The uncertainty representation and counterfactual mechanisms are essential for scientific rigor—every label is traceable to a mechanistically explicit model and a set of plausible alternative worlds established by executable do-interventions.

4. Image-to-Simulation Pipeline: An Operational Example

The V&V instrumented image-to-simulation ("img2sim") pipeline demonstrates a full realization as follows:

1. Input: A single photograph of a physical artifact (e.g., loaded metal bracket).
2. Perception: Multi-agent LLM system extracts geometry and material properties, establishing $I_i$8, $I_i$9, $\Omega_i$0, etc., with uncertainty estimates.
3. Model building: Automatic meshing and solver set-up ($\Omega_i$1).
4. Verification: Mesh convergence, residual checks with respect to analytical bounds.
5. Solution: Solver produces field ($\Omega_i$2).
6. Postprocessing: Quantity of interest $\Omega_i$3 is derived.
7. Validation: Compilation of a domain-expert V&V report ($\Omega_i$4).

The pipeline iteratively produces $\Omega_i$5, an uncertainty cloud $\Omega_i$6, and auditable counterfactual families by directly manipulating model or confounder parameters (e.g., load, material stiffness, illumination), with complete V&V traceability for all variants (Wilke, 5 Jun 2026).

5. Uncertainty Quantification and Decomposition

Instrumented Data represent uncertainty at every modeling stage. Each parameter $\Omega_i$7 comes with:

• $\Omega_i$8: Extraction interval.
• $\Omega_i$9: Density estimate over parameter values.

The push-forward through model, solver, and post-processing yields $\sigma_i$0, a distribution over quantities of interest, representing the observational uncertainty structure. Using repeated extraction and calibration (potentially human-in-the-loop or conformal), uncertainties may be decomposed into:

• Aleatoric ($\sigma_i$1): Irreducible uncertainty due to inherent system randomness.
• Epistemic ($\sigma_i$2): Reducible uncertainty, tracing to imperfect knowledge or extraction limitations.

The operational protocol for this decomposition remains an open question in current research.

6. Programmatic Construction and Auditing

Instrumented Data are constructed programmatically with explicit logging and traceability at each step. The workflow (illustrated in (Wilke, 5 Jun 2026)) is:

$\sigma_i$7

Graphically, the data flow is $\sigma_i$3, with $\sigma_i$4 parallel to $\sigma_i$5, uncertainty $\sigma_i$6 branching, and parallel tracks spawned by do-interventions.

7. Implications and Future Directions

Instrumented Data establish a substrate for scientifically grounded machine learning, enabling:

• Mechanistic supervision in surrogate modeling and transfer learning.
• Systematic, auditable causal interventions.
• Domain-specific V&V, ensuring scientific validity and generalizability.

Immediate applications span computational biology, climate modeling, materials science, fluid mechanics, and medical imaging, where observational and synthetic data alone are insufficient for causal inference or robust supervision. A plausible implication is that, in the long term, instrumented data may underlie the construction of foundation models for scientific reasoning, providing a falsifiable and mechanistically interpretable substrate for automated discovery and hypothesis testing (Wilke, 5 Jun 2026).