GoKawiilMultimodel ensemble approach
We simulated scenarios using a MME, building on the global economics group of AgMIP food system modelling efforts that have developed modelling protocols to explore future uncertainty and increase the comparability across various economic models13,16,21,59. The ensemble includes ten global economic models, which have featured in a wide range of high-level science policy reports (for example, ref. 60). The participating models are AIM, CAPRI, ENVISAGE, FARM, GCAM, GLOBIOM, IMAGE, IMPACT, MAGNET and MAgPIE. Although all models are global in scope, they have varying levels of detail in representing parts of global food systems, including food production, processing, trade and demand, as well as environmental and socioeconomic outcomes of these activities (Supplementary Tables 3 and 4 give a summary of participating models).
Scenarios were developed by the MME coordinators and the individual model teams. The main scenario (EL2) was developed to mirror the core components in the 2025 EAT–Lancet report—shift to a healthy reference diet, increased agricultural productivity and a reduction in FLW, with the BAU scenario as a counterfactual to EL2. The scenario specification was shared with modelling teams to implement in their models. To align models on key inputs, data on the healthy reference diet, gross domestic product (GDP) and population projections from SSP2 v.3, productivity trends and climate shocks to crops, livestock and labour were shared with model teams. Model teams ran scenarios and shared results with the MME coordinators who processed and combined individual model scenario submissions. To ensure comparability, model outputs were converted to relative changes against 2020. There was a continual process of feedback and iteration during the exercise to harmonize on how models implemented scenarios and to compare preliminary results. The main forum for this was a monthly virtual meeting between all project participants supported by more frequent dedicated one-to-one engagement between the MME coordinators and individual model teams. In total, there were six main submission rounds—(1) initial ‘zero’ round run to test the process, (2) integrate v.3 of SSP2, (3) harmonize on BAU climate policies, (4) integrate updated healthy reference diet, (5) revise agricultural productivity target and (6) FLW sensitivity analysis—that were accompanied by individual model iterative submissions from modelling teams between October 2023 and July 2025. Figure 5 summarizes this process.
Fig. 5: MME process undertaken for this study. Full size image Key data inputs and the main modelling results submission rounds are detailed.
There were ten participating models in the study. Each has somewhat different structures in how they represent food, environmental, biophysical and economic systems. For detailed documentation on how each of these models captures these systems, behaviours and feedbacks, please follow the links in Supplementary Table 3 and see Supplementary Information for summaries.
Scenarios and data inputs
Five scenarios and a further four sensitivity scenarios are presented here (Supplementary Table 1). Scenarios make use of the latest SSP2 data5 for population and GDP projections. We opt for SSP2 as it represents a ‘middle of the road’ pathway (see Supplementary Fig. 1 and Supplementary Table 2 for further details on SSPs). The first is a BAU scenario that reflects current trends based on SSP2 projections for key scenario drivers (GDP and population) and assumptions related to diets, crop and livestock productivity and FLW. The food systems transformation scenario (EL2) also uses SSP2 GDP and population projections but includes the EAT–Lancet reference diet, along with higher productivity growth rates similar to trends associated with SSP1 (SSP1 represents a more optimistic ‘sustainability’ path) and a reduction of FLW broadly consistent with the 50% reduction contained in Sustainable Development Goal 12.3. The final three scenarios represent individual drivers of this food systems transformation (shifts to the EAT–Lancet diet (BAU_DIET), improved productivity (BAU_PROD), and reduced food loss and waste (BAU_WAST)). Four sensitivity scenarios explore FLW reduction of 25% and 75% in the BAU and EL2 core scenarios, respectively.
The diet implementation requires data on food consumption categorized by food group and region for both the BAU and EL2 diet, provided by refs. 2,18. Supplementary Table 8 shows the global average reference intake values to be achieved by 2050. Models assume that populations make changes to the reference diet by way of a consumer preference shift. We assume that EAT–Lancet healthy reference diets will be achieved in all regions by 2050. Caloric coefficients (cal g−1) are held constant in the projection period. For food groups such as fruits and vegetables the intake value (g per day) was interpreted as a floor and populations in the models were free to consume more than this quantity. By contrast, many of the animal-sourced food groups were interpreted as a ceiling in which populations above this converged down to the target, but populations below this continued to follow country or region dietary trends up to and including this value.
The approach to the agricultural productivity scenario component was as follows. Future trajectories of potential growth in agricultural sector productivity have been developed by researchers at the International Food Policy Research Institute (which hosts the IMPACT model) over the last two decades61. Yield growth rate assumptions in IMPACT are periodically updated through consultation with experts (for example, Consultative Group for International Agricultural Research (CGIAR) centres), economic model comparison projects, trends in agricultural research expenditures and updates on trends in long term yield growth rates based on FAOSTAT data62. Yield trends from this process formed the BAU productivity growth rates. The higher rates in EL2 were constructed by scaling BAU rates by the difference in per capita GDP between SSP1 and SSP2 (see refs. 63,64 for details). On average, this gave a 10–15% percentage point increase in yield between BAU and EL2 over 2020–2050 (Supplementary Table 9). Yields in EL2 and BAU versus historic trends are given in Supplementary Fig. 11.
Scenarios did not include any more mitigation policies that have not already been implemented. Some specific climate-related impacts are included in line with representative concentration pathway (RCP) 7.0. We opt for RCP 7.0 as it broadly reflects the outcome of no further climate policy, representing an upper range of future GHG emissions and warming by 2050. These include climate-related impact shocks to crops (using IPSL GCM average of the global gridded crop model intercomparison (GGCMI) crop model ensemble for soybean, maize, wheat and rice65), livestock66 and agricultural labour67. The four main crops represented by the global gridded crop models (maize, wheat, rice and soybeans) were mapped to 36 crop commodities of IMPACT as closely as possible to their native crop representation. Likewise, the livestock climate shocks for meat and dairy were mapped to beef, lamb, pork, poultry, eggs and milk, respectively. Agricultural labour productivity trends are based on ref. 67 and reflect global reductions in manual agricultural work capacity owing to climate change. Supplementary Table 5 details how individual models implemented the scenario components.
... continue reading