R&D stage · real public-data benchmark

Digital Field Twin

A physics-informed decision engine for crop genetics-×-environment trial analytics — benchmarked on real public research data and gated so it abstains rather than guesses when the evidence isn't there. The decision layer that tells you when it can't predict.

What it is

Breeding programs must decide which untested entries to advance and which site-years run high or low risk — but field phenotyping is expensive and sparse, so most candidate genotypes and environments are never directly measured. The Digital Field Twin composes four evidence-gated modules — Genetics (genomic prediction of untested germplasm), Weather/Environment (environment yield level and stress context), Imagery/EO (satellite & UAV yield signal), and Agronomy (treatment response) — over real public datasets: Genomes-to-Fields maize, SoyNAM soybean, CIMMYT wheat, NASA POWER weather, and more.

Every claim is consolidated into a machine-generated scorecard with an explicit confidence type and a plain-English decision meaning — and a selective-prediction layer that declines to predict outside its support.

Why it's different

Breeding platforms (Phenome Networks), imagery vendors (Hiphen, Solvi, Sentera), and yield-forecast engines (Cropin, Pattern Ag) all sell a confident predictor or a data-management system. None publicly ships an evidence-gated scorecard that publishes its own negative results, or an explicit abstention layer. The Digital Field Twin does both — it separates value by axis (genetics ranks untested germplasm; weather sets the environment level; they compose rather than compete) and reports the 14 things that don't work alongside the 14 that do.

What we don't claim. This is R&D-stage decision support on public benchmarks — not a live, real-time "digital twin" of your field, and not validated on customer data. Genetics-forward prediction of genuinely new hybrids is not supported (a data/relatedness ceiling, shown with CI-backed negatives). Treatment-response findings are associations, not causal. We lead with where the evidence is strong and name where it isn't.

Prepared demo

The evidence scorecard — negatives included.

Every product claim carries a status and a confidence type. Publishing the not-supported findings alongside the supported ones is the point, not an omission.

Supported14
Conditional12
Not supported14
Gap1

41 findings from leakage-controlled splits on public datasets. A distribution of evidence, not a percent-complete score — a new module can add both supported and not-supported rows while genuinely improving the system.

Start here

Run a Measurement Validation Sprint on your trial data.

A fixed-scope, 2–3 week engagement on your archived trial data with a pre-registered acceptance test and an auditable go / no-go. The honest way to find out if this helps your program.

Project inquiry

For seed / breeding R&D and analytics teams.

Email ASI Labs