Research prototype

Generative concrete mix screening.

Generate and screen many mix candidates with the same acceptance logic.

Supplied research-workflow diagram; a diagram of the study, not a live application screenshot.

Role: Research author · credited application developer
Period: 2026
Methods: Flow matching · Python
Status: Research prototype

Engineering software / Streamlit

A defined input.
An inspectable decision.

For concrete materials and computational engineers

Generate and screen many mix candidates with the same acceptance logic.

01 / INPUT

Set the engineering case

28-day strength target, water/aggregate settings and screening budget

02 / OUTPUT

Inspect the result

Mix candidates with feasibility, domain and uncertainty diagnostics

03 / DECISION

Compare the next options

Choose candidates for review and physical testing.

Technical basis

Flow matching · Python

Validation evidence

Credited Stage09 source, inference engine and frozen-bundle documentation.

Current scope

Implemented research workflow; the public visual is its workflow diagram, not a running application screenshot.

Public deployment has not been verified. The supplied interface and source establish the implementation shown.

Engineering record

Scope, method and evidence.

Implemented Stage09 research workflow; laboratory qualification and production deployment remain outside the evidence.

01A mixture candidate is a starting point for verification

A strength target can correspond to many concrete mixtures. The implemented Stage09 research tool generates candidate compositions and screens their predicted response, rather than claiming a unique or laboratory-qualified mix.

The primary target is 28-day compressive strength. The page follows the supplied Streamlit application and engine lineage; it does not merge capabilities from a separate proposed web architecture.

02Implemented source and project status

Application source, an inference engine and a frozen-bundle README support an implemented Stage09 workflow. The Stage09 source explicitly credits Sina Sarfarazi as the application developer.

This credit does not imply sole authorship of underlying data or every model component. Deployment instructions do not demonstrate a running public release.

03What the implemented pipeline does

Generate compositional alternatives
Predict one of 16 ingredient-activity regimes, obtain a conditional compositional centre and use masked residual conditional flow matching for one-to-many candidate generation.
Reconstruct physical composition
Preserve structural zeros and enforce binder closure in reconstruction. These are representation checks, not proof of fresh or hardened concrete performance.
Screen with forward models
Random Forest, Extra Trees, HistGradientBoosting and XGBoost models provide strength predictions. Descriptor support and ensemble disagreement are exposed as diagnostics.
Rank within a declared scenario
Use empirical q80/q90 calibration diagnostics and material-stage GWP/reference-cost Pareto comparisons to inspect retained alternatives.

04Available implementation evidence

The supplied README, application and engine define the model-loading, reconstruction, screening and reporting sequence. No live demonstration is linked because a current deployment has not been verified.

The supplied research-workflow diagram appears above. It is not a running Stage09 screenshot.

Scroll the table horizontally to inspect all columns.

Meaning of the reported diagnostics
Diagnostic	Supported interpretation	Not established
q80 / q90 thresholds	Empirical calibration diagnostics	Formal conformal coverage guarantee.
Forward strength predictions	Surrogate-based candidate screening	Laboratory validation.
Material-stage GWP	Cradle-to-gate comparison using generic literature factors	Complete project LCA or verified carbon reduction.
Reference cost	Declared material-cost scenario	Current local quote or life-cycle saving.

05Research boundary

Workability, durability, exposure class, chloride resistance, shrinkage, creep, early-age requirements and code-specific qualification are outside the primary prediction target. These need separate evidence and checks.

Cross-study transfer is harder than pooled in-domain prediction. A new material system may require additional descriptors or local calibration. No generic “accuracy”, compliant mix or delivered field benefit is claimed.

Manuscript metrics are not used as headline results until they are reconciled with the implemented release and evaluation split. Software licensing and public release status also need confirmation before distributing the research source.

06Related engineering software

Engineering R&D explains how a calculation or screening workflow can be scoped for a new assignment.

Discuss a decision workflow