Modelling complex agrifood processes like for instance cheese ecosystems is challenging due to the complex multi-scale nature of interactions among and between the constituent components. Furthermore, producing experimental data in this context is both tedious and expensive, resulting in scarce datasets, where the number of dimensions far exceeds the number of data samples. We present here LiDeoGraM, a visual analytics tool that combines data-driven machine learning with domain experts’ knowledge to produce multi-scale models of living ecosystems. This tool is inspired from the idea of model stacking where an ensemble of simple local models are generated for each component of the studied system. The modelling process is carried out in three iterative steps: (a) using regression, LiDeoGraM proposes a set of local models for each component; (b) via a graphical interface, domain experts evaluate those local models; (c) an interactive evolutionary algorithm builds a global model while mediating between expert’s subjective assessment, and an automatic evaluation based on fitting error and complexity. A first validation of our approach has been performed at three levels, combining computational and human-centered evaluations: (i) automatic tests to assess the robustness of the local model generation; (ii) a toy model test to evaluate how domain experts use our tool to discover a ‘known’ model; and (iii) a use case study to examine how domain experts use LiDeoGraM to model real-life multi-scale biological systems. Our results show that domain experts are able to operate our tool to discover a known model, and are able to generate new hypotheses when exploring their own datasets.