When a property is missing, an applicable model can suggest what to expect. Every estimate comes with its inputs, benchmark, limits and a way to inspect the supporting calculations.
Prototype composition models
These models estimate the median of published DFT calculations for a composition. Crystal structure, processing, defects and operating conditions can change the real material. Their outputs are proposals to test.
Try a property estimate
Explore this held-out example freely. Sign in with a CogLab membership to evaluate your own compositions.
K3Cu3As2 · electronic band gap · held-out example
Estimated · requires verification
0.723 eV
Calibrated interval: 0–1.791 eV
Withheld JARVIS target: 1.363 eV. This is a computed reference for comparison, not a measurement.
What informed this estimate?
K2CuAs · DFT median 1.282 eV
Composition distance 0.25 · provider structure range 1.282–1.282 eV
Target is the median of JARVIS DFT values for a reduced composition, not a measured property or a structure-specific prediction.
Crystal structure, polymorph, defects, microstructure, temperature and pressure are not inputs.
Only the same chemical system with three nearby training compositions is eligible. No claim of generalization to new chemical systems.
Random composition-group split prevents identical reduced compositions crossing partitions. It does not establish a prospective experimental benchmark.
The interval is calibrated on held-out eligible compositions from this dataset. It is not an experimental uncertainty or a per-material probability guarantee.
Reduced compositions are grouped before a fixed 60/20/20 training/calibration/test split. The gate requires 30 eligible calibration and test compositions, improvement over the baseline, and at least 80% observed coverage. No parameters were tuned on these tests. This evaluates a narrow dataset subset, not experimental accuracy.
Reduced compositions are grouped before a fixed 60/20/20 training/calibration/test split. The gate requires 30 eligible calibration and test compositions, improvement over the baseline, and at least 80% observed coverage. No parameters were tuned on these tests. This evaluates a narrow dataset subset, not experimental accuracy.
Use the decision engine for unit-aware requirements, uniform slab transport and exact reaction conservation. Use lab analysis for supported measurement workflows. Deterministic calculations validate their stated model; they do not turn assumed inputs into measured properties.