Why Modeling Is Critical for Resilient Agricultural Production Systems

By: Bruno Basso, Ph.D.‍
‍John A. Hannah Distinguished Professor, Michigan State University | Co-Founder, Terion AI | AAAS Fellow‍

‍
Every consumer-packaged goods (CPG) company that has made a climate or regenerative agriculture commitment now faces the same uncomfortable arithmetic: most of its footprint sits in farm fields it does not own, cannot visit, and cannot measure directly. A cereal brand sourcing oats from thousands of farms, a brewer buying barley across three continents, a food company pledging lower-carbon dairy – all of them need to know what is happening in the soil, under different weather, on land they will never sample. Nor is this challenge unique to CPGs; any organization making decisions about agricultural outcomes faces the same need for field insights. Only one scientific instrument can turn those insights into reliable estimates at that scale, before the money is spent and the claims are made: agricultural systems modeling.

Crop-soil models simulate conditions in the soil and the plant under different management strategies and genetics on a computer, so errors can be made and corrected in the digital world instead of the real one, where they would cost money and resources. A farmer, or a company sourcing from ten thousand farmers, can learn a crop’s footprint in an environment before that crop is even planted. That is the true power of system-based modeling, and it is the foundation on which credible sustainability programs, and, more broadly, a more abundant agricultural ecosystem, are now being built.

Sixty years of history

Agricultural modeling is sometimes presented as a new technology, but it is not, and its long history is precisely why it can be trusted. As colleagues and I documented in a review of the field’s evolution (Jones et al., 2017), the foundations were laid in the 1950s and 1960s, when pioneers such as C.T. de Wit in the Netherlands began the first computational analyses of plant and soil processes – asking, with the primitive computers of the day (punch cards), whether photosynthesis, water movement, and crop growth could be written down as equations and solved.

A geopolitical shock provided the decisive push. In 1972, the Soviet Union quietly purchased enormous quantities of U.S. wheat, catching the government by surprise, driving up world prices, and exposing how little anyone knew about global crop production in real time. Federal agencies funded new research programs to build models that, combined with newly available satellite observations, could estimate the production of major crops anywhere in the world (AGRISTARS and IBSNAT projects). Out of that effort came the CERES family of wheat and maize models developed byJoe Ritchie and colleagues – models whose descendants are still in use today, and whose lineage runs directly into SALUS, the model my group developed at Michigan State University to simulate whole cropping systems, continuously, over decades, crops, soil water, nitrogen, carbon, and management, together, over space and time.

The international IBSNAT project of the 1980s carried models into the developing world; the first IPCC assessment in 1990 made them the standard tool for climate-impact analysis; economic and livestock models joined crop models to represent farms and food systems rather than single fields; and in 2010 the launch of AgMIP, the Agricultural Model Intercomparison and Improvement Project (Rosenzweig et al.,2018) organized the world’s modeling groups into a single coordinated community. Sixty years of testing, comparison, failure, and refinement is the pedigree behind every simulation run today. Very few technologies deployed incorporate sustainability can claim the same depth of scientific vetting.

Why models: agriculture is a system, not a list of practices

The reason crop-soil models are indispensable – and the reason agricultural practice checklists and emission factors keep disappointing – is that agriculture is a system, a complex system. Yield, carbon, nitrogen, and water outcomes emerge from the interaction of genetics, environment, and management: G × E × M, played out through the soil. The same cover crop that builds carbon on loamy soil may cost yield on cold clay; the same nitrogen rate that is optimal in a wet year is wasteful and polluting in a dry one; the same field contains zones that behave differently every season. A process-based model represents the mechanisms, radiation capture, soil water balance, nitrogen transformations, carbon turnover, and lets them interact, day by day, under the actual weather and soil of each place.

This is what makes a model a “digital twin” of a field. On the screen, management can be changed and the consequences observed before they happen in the ground. “What If” I change my tillage from conventional tillage to no-till, how much do my GHG emissions change? “What if” I plant a cover crop, how much carbon do I accrue on this soil, in this climate, under this management? Can my soil supply water to roots when the crop needs it, or will water be lost to groundwater, taking nitrate with it? Simulating a scenario before it happens serves short-term decisions, and because the weather inputs can be replaced with future climate projections, it is also the only way to test how a sourcing region, a rotation, or a regenerative program will perform in the climate of 2040. For any organization sponsoring, regulating, or depending on agricultural outcomes, that means baselines, intervention design, and supply resilience can all be evaluated with one coherent scientific instrument instead of a patchwork of factors and assumptions.

The lesson that changes everything: there is no single best model

Here is the finding that two decades of model intercomparison has made unavoidable, and that separates scientifically mature programs from the rest: at scale, there is no single best model. Run several well-tested models across many sites and years and a consistent pattern emerges, one model is more accurate in one location and one season; another wins somewhere else, in another year. Model A captures the dry-year water dynamics of a sandy soil; Model B better represents carbon turnover on a poorly drained clay. Each model is a defensible, incomplete hypothesis about how the system works, and reality takes turns agreeing with each of them. This creates a temptation and an opportunity. The temptation is cherry picking: with enough models available, one can always find the single model whose answer flatters the program, the highest carbon accrual, the largest footprint reduction, and report it. Nothing in the output betrays the selection. The opportunity is the opposite move: use all the credible models together, as a multi-model ensemble (MME), and let their agreement and disagreement become information.

The logic is the sameone that economists have used for half a century in forecast combination – the“pooling” of forecasts pioneered by Bates and Granger in 1969. Individual forecasters, like individual models, carry systematic biases; combining them cancels errors that no single forecast can cancel for itself. In agricultural modeling, the evidence is now overwhelming: across the AgMIP intercomparisons for wheat, maize, and rice, the ensemble median predicted observations better than any individual model – including the best one, which could not have been identified in advance. And the ensemble delivers something a single model never can: the spread across models is an honest, quantitative measure of how certain the science is, place by place and practice by practice.

In our recent work formalizing MME frameworks for agricultural sustainability (Basso et al., 2025), we take this a step further. An ensemble is not a black box that averages blindly; it is a structured synthesis of models that have each been independently tested, in which the ensemble mean and median become new, stronger estimates precisely because they inherit the accumulated validation of every member. And those central estimates are then themselves subjected to a further test – domain knowledge. Does the simulated carbon accrual saturate the way long-term experiments say it may? Does the nitrogen response respect what forty years of field trials established? When soil carbon or emission reductions are quantified for credits or Scope 3 claims, the defensible number is not the most optimistic model’s output; it is the ensemble’s conservative bound, reported with its range. This is the principle now advancing in international accounting discussions, from soil carbon methodologies under aviation’s CORSIA framework to the EU’s carbon removal certification: credit what the ensemble supports, disclose the uncertainty, and let re-measurement close the gap. Programs built this way survive scrutiny. Programs built on a single hand-picked model increasingly will not.

Why Terion AI

This is the frontier on which Terion AI was deliberately built. Terion operates a true multi-model ensemble: not one model with a dashboard, but multiple, independently validated process-based models that run together, their agreement quantified, their spread reported. This MME is coupled with the deep domain knowledge base of a team that has spent decades developing, testing, and publishing these models across soils, climates, and cropping systems worldwide. That combination matters because an ensemble without domain expertise cannot be interrogated, and domain expertise without an ensemble cannot be scaled. Terion has both: the machinery to generate scientifically defensible, uncertainty-aware estimatesfor every field in a supply shed, and the human knowledge to test every ensemble result against what agricultural science has established – the modern, computational equivalent of asking a panel of the world’s best agronomists and taking their considered consensus rather than the loudest voice. The result is independent, accurate, and reliable intelligence you can act on with confidence.

For organizations making decisions about agriculture, the practical translation is this: baselines that hold up in audit; interventions chosen for the fields where models agree the response will be largest; environmental claims quantified with the conservatism that registries, regulators, and increasingly consumers demand; and a transparent measure, the ensemble spread, of where confidence is high and where ground measurement should be invested. Sixty years ago, a handful of scientists began writing the growth of a plant as equations. Today, the responsible use of that inheritance is not to pick the equation we like best. It is to pool everything the models have learned, test it against everything we know, and act on the consensus. That is what modeling is for, and it is how agriculture’s regenerative transition will be measured, financed, and trusted. This is why modeling is critical for resilient agricultural production systems.

‍

Bruno Basso is a John A. Hannah Distinguished Professor and MSU Research Foundation Distinguished Professor at Michigan State University, a Fellow of the American Association for the Advancement of Science, cofounder of CIBO Technologies, and Terion AI and leads Terion AI Labs. His research on crop and soil modeling, yield stability, and agricultural sustainability spans field to continental scales, from smallholder farms in developing economies to precision agriculture across the US Midwest. He is an AAAS Fellow.

‍

‍