Structural Sculpting: Making Inverse Modelling Generate and Deal with Variable Structures

When dealing with Agent-Based Models (ABMs), calibration, sensitivity analysis, and robustness testing are often limited to parameter space and seeding, while structural calibration is omitted. However, we know that model structure necessarily also influences model outcome. Omitting structural calibration would thus pose a significant hurdle to robust model-based decision support, policy evaluation, and behavioural insights. Inverse modelling is an explorative modelling approach newly introduced for ABMs, aimed at directly inferring the generative mechanisms underlying observed outcomes by iteratively posing forward problems to match the ABM output with the desired patterns. We propose a method that leverages the inverse method on an ABM's building blocks to calibrate the model for generative insights structurally. We exemplify this through a case study using a solar panel diffusion model with Dutch province-level data, for which we operationalise "structure" through the order and presence or absence of procedures called in the model iteration. Our method shows that it is possible to vary and evaluate model structures automatically via inverse modelling. We find structures that fit each province’s solar panel adoption curve well and others poorly, and that variations, structural or in seed, significantly influence model outcome. We find multiple alternative well-performing model structures that exhibit large deviations concerning order and even the presence of functions. We exemplify how these structures can be made sense of and point directions for further real-life investigations and theory-building, such as the effect of hassle factors or complexity perceptions on adoption rates. With this, we present not an automated replacement of the participatory modelling process but an add-on to systematically reflect on the structure, implementation, and validity of the ABM and the theory utilised.