Crop simulation models of plant processes capture the biological interactions between the sensing of signals at an organ level (e.g. drought affecting roots), the response of the plant at a biochemical level (e.g. change in development rate) and the result at the organ (or crop) level (e.g. reduced growth). In dissecting the complex control of phenotypes like yield, simulation models have several roles. Models have been used to generate an index of the climatic environment (e.g. of drought stress) for breeding programme trials. In wheat and sorghum grown in northern Australia, this has shown that mid-season drought generates large genotype by environment interaction. By defining gene action to calculate the value of input trait parameters to crop models, simulated multi-environment trials estimate the yield of ‘synthetic’ sorghum cultivars grown in historical or artificial climates with current or potential management regimes. In this way, the biological interactions among traits constrain the crop yields to only those that are biologically possible in the given set of environments. This allows the construction of datasets that are more ‘realistic’ representations of gene by trait by environment interaction than is possible using only the statistical attributes (e.g. means, variances and correlations) of real-world trait datasets. This approach has an additional advantage in that ‘biological and experimental noise’ can be manipulated separately. These ‘testbeds’ for statistical techniques can be extended to the interpretation of a crossing and selection programme where the processes of chromosomal recombination are simulated using a quantitative genetics model and applied to the trait parameters. Statisticians are challenged to develop improved methods for the resulting simulated phenotype datasets, with the objective of revealing the (known) underlying genetic and environment structure that was input to the simulations. These improved methods can then be applied to existing plant breeding programmes.