AI-assisted engineering inference for simulation-driven workflows.

Research direction

SIMAIS investigates workflows in which established engineering information and explicit computation can be combined with AI-assisted inference to support estimation and evidence-driven decisions.

The broader goal under investigation is whether the following can be combined to support engineering estimation and evidence-driven decision workflows:

• engineering geometry or system descriptions
• operating conditions
• existing reference data
• AI-assisted inference
• explicit numerical methods

Broader engineering applicability beyond the current experiment remains a research hypothesis. That generalization has not been demonstrated.

What we are building

SIMAIS is investigating a reusable engineering inference workflow that combines:

• engineering geometry or system descriptions,
• operating conditions,
• existing reference evidence,
• AI-assisted inference,
• explicit numerical methods,

to produce fast, evidence-backed engineering estimates while keeping numerical computation auditable.

The current airfoil Cd experiment is the first testbed for this workflow, not the product boundary.

This broader capability has not been demonstrated. It remains a research hypothesis under investigation.

Intended users

Prospectively, SIMAIS is intended for audiences such as:

• engineering students,
• researchers,
• small engineering teams exploring simulation-driven design or estimation workflows.

These are intended users only. SIMAIS has no existing customers and no production deployment is claimed.

Current experimental domain

The currently implemented and measured pipeline focuses on two-dimensional airfoils at Reynolds number 100,000 and angle of attack 0°.

This is the first experimental testbed for SIMAIS. It is not the intended permanent scope of the project, and SIMAIS is not an airfoil-only, CFD-only, or aerodynamics-only project.

It combines:

• public airfoil geometry
• archived numerical Cd references
• geometry-based reference retrieval
• a hosted external decision model called Jev
• an explicit calibrated numerical decoder
• auditable aggregation
• comparison against archived CFD references

Jev does not perform CFD, and the system does not generate CFD solutions. Reported comparisons are against archived numerical references, not newly computed flow fields.

How the current prototype works

The prototype below describes the measured airfoil Cd estimation path. It is intended to be understandable to an engineering audience.

1. Ingest public airfoil geometry and numerical Cd references.
2. Organize available geometry/reference data for experimental roles.
3. Retrieve geometrically similar support profiles with known Cd.
4. Submit query/reference comparisons to Jev.
5. Validate the returned probability distributions.
6. Convert those probabilities into anchor-relative Cd estimates using an external calibrated numerical decoder.
7. Aggregate valid estimates in log-Cd space.
8. Compare the resulting estimate against the archived numerical reference.

Development benchmark — not independent validation

The values below summarize exposed, adaptive development comparisons on a fixed assessment set. They describe the current prototype only.

Caveat: These are exposed, adaptive DEVELOPMENT comparisons against archived finite-iteration CFD references. They are not independent validation, do not demonstrate 2% physical accuracy, and do not establish that the system can replace CFD or established engineering methods.

All three assessment cases with reference Cd above 0.03 missed the 2% threshold.

Two later same-method runs produced 101 and 102 cases within the same 2% criterion on the same assessment set. Those repeated runs are reported here for transparency and are not independent replication.

Evidence and limitations

The reported development comparisons should be read with the following limitations:

• the references are finite-iteration CFD outputs;
• convergence remains qualified;
• exact geometry/solver correspondence remains qualified;
• turbulence-model lineage remains qualified;
• the arithmetic used to reproduce predictions was independently audited;
• the original assessment wrappers did not fully enforce label-access isolation before the experiment was sealed;
• Jev’s isolated contribution has not been established.

Research boundaries

The following remain unestablished:

• independent performance on genuinely new geometry groups
• generalization beyond the current Reynolds number and angle-of-attack condition
• calibrated Cd uncertainty
• successful CFD-specific native model adaptation
• superiority over established engineering methods
• CFD replacement capability
• production readiness
• commercial validation

No customers or production deployment are claimed.

Broader engineering applicability remains a research hypothesis.

Role of AI and numerical methods

Jev is an external decision model provided by TypeSafe AI.

In the tested prototype, hosted jev-1.13.0 returns probabilities over seven relative-drag categories for query/reference comparisons.

• Jev’s model weights are not modified.
• A separately calibrated numerical decoder converts those probability outputs into Cd estimates.
• Numerical aggregation and scoring remain explicit and auditable.

Jev is not a CFD solver, a physics simulator, a CFD replacement, or the full SIMAIS architecture.

Model documentation: https://docs.typesafe.ai/models

Planned Claude layer

This section is prospective. It describes work under evaluation, not the measured prototype.

The intended role is not to replace numerical solvers. Claude is being evaluated as an evidence and workflow reasoning layer that may help connect heterogeneous engineering information without silently replacing explicit numerical computation.

Claude is being evaluated as a future reasoning and workflow layer that may help:

• interpret heterogeneous engineering evidence
• detect differences in operating conditions
• track dataset provenance
• organize experiments
• turn auditable numerical results into structured engineering reports and decisions

Claude API is not part of the measured Cd prediction prototype described on this page.

The numerical prediction path should remain explicit and auditable rather than asking a generative model to silently replace numerical engineering code. Claude is not currently producing the reported Cd estimates.

Data

The primary public dataset for the retained experiment is:

A Comprehensive Dataset of the Aerodynamic and Geometric Coefficients of Airfoils in the Public Domain
Data 2024, 9(5), 64

• DOI: 10.3390/data9050064
• Paper: https://www.mdpi.com/2306-5729/9/5/64
• Upstream dataset: kanakaero/Dataset-of-Aerodynamic-and-Geometric-Coefficients-of-Airfoils

This dataset supplies the geometry representations and aerodynamic reference data underlying the reported retained experiment.

UniFoil, NREL Airfoil2k, and AFBench were only investigated as possible additional data sources. They did not contribute to the reported 101/160 result.

Project status

Early-stage R&D prototype.

Current work focuses on:

• evidence quality
• broader usable datasets
• independent testing
• determining whether the method generalizes beyond the initial airfoil experiment

Model learning / CFD-specific adaptation has not yet been established as successful.

This project is not production-ready.

Contact

For research inquiries:

founder@simais.dev