💻 computer science

Agent-Based Modeling of Low-Emission Fertilizer Adoption for Dairy Farm Decarbonisation using Empirical Farm Data

This study develops and validates an agent-based modeling framework using empirical data from 295 Irish dairy farms to simulate the socio-technical diffusion of low-emission fertilizers over 15 years, demonstrating how social networks and policy interventions drive adoption trajectories and offering a robust tool for evaluating decarbonization strategies.

Original authors: Surya Jayakumar, Kieran Sullivan, John McLaughlin, Christine OMeara, Indrakshi Dey

Published 2026-05-06

📖 4 min read☕ Coffee break read

CC0 1.0

Original authors: Surya Jayakumar, Kieran Sullivan, John McLaughlin, Christine OMeara, Indrakshi Dey

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a giant, high-tech simulation game where the "players" are 295 real dairy farms in Ireland. The goal of the game isn't to win points, but to figure out how to stop these farms from polluting the air with greenhouse gases.

The researchers built a digital twin of the Irish dairy industry to test a specific idea: What happens if farmers switch from regular fertilizer to a "low-emission" version (called Protected Urea)?

Here is the story of their findings, broken down into simple concepts:

1. The Problem: Not All Farms Are Alike

Think of the dairy industry like a classroom. Some students are "A-players" (benchmark farms) who get great grades (high milk production) with very little effort (low pollution). Others are "C-students" (laggard farms) who struggle, use too much fertilizer, and create a lot of pollution for the amount of milk they produce.

The researchers realized that you can't treat every farm the same. A "one-size-fits-all" rule won't work because the reasons for pollution are different for everyone. Some farms are just inefficient; others are struggling with too many cows on too little land.

2. The Engine: The "Social Network" Game

The most interesting part of this study is how they modeled how farmers make decisions. They didn't just assume farmers look at a price tag and decide. Instead, they built a social network into the computer model.

The Analogy: Imagine a small town where everyone talks to their neighbors. If your neighbor starts using a new, better fertilizer and tells you, "Hey, it works great and saves me money," you are much more likely to try it too.
The Finding: The computer showed that farmers are heavily influenced by their peers. If a few early adopters switch, their neighbors follow. This creates a "snowball effect." The model proved that if you ignore this "gossip" or peer pressure, your predictions are wrong. The social network is the engine that drives the change.

3. The Experiment: Subsidies vs. Taxes

The researchers used their simulation to test two different ways to get farmers to switch:

Scenario A: The Carbon Tax (The Stick)
Imagine the government puts a fine on polluting fertilizer. It makes the dirty stuff expensive.
- Result: It works, but slowly. Farmers grumble and switch eventually, but it takes a long time.
Scenario B: The Subsidy (The Carrot)
Imagine the government pays farmers a bonus if they buy the clean fertilizer.
- Result: This worked much faster. Because the clean option became cheaper, the "early adopters" switched immediately. Once they switched, their neighbors saw them doing it and followed suit. The subsidy jump-started the social snowball.

The Verdict: The subsidy was the winner. It got the job done faster and resulted in more total pollution reduction over 15 years because the change happened sooner.

4. The "Policy Laboratory"

The authors call their computer model a "Policy Laboratory."

Think of it like a flight simulator for pilots. Before a real plane takes off, the pilot practices in a simulator to see what happens if the engine fails or the weather turns bad. Similarly, before the Irish government spends real money or passes new laws, they can run this simulation to see:

Will farmers actually do it?
How fast will it happen?
How much money will it save or cost?

The simulation showed that with the right mix of money (subsidies) and peer pressure, the industry could switch to clean fertilizer with a 91% success rate. The remaining 9% of farmers are the "laggards" who just won't switch, no matter what you do.

5. The Bottom Line

It's Cheaper Than You Think: Switching to the clean fertilizer is actually a bargain. The cost to the farmer to reduce one ton of pollution is very low (less than €1), which is much cheaper than current carbon prices.
Social Pressure is Key: You can't just force farmers to change with rules; you have to help them see their neighbors doing it first.
The Tool Works: The computer model was so accurate (97.9% accuracy) that it could predict real-world behavior almost perfectly.

In short, the paper says: "If you want to clean up Irish dairy farming, don't just punish the polluters. Pay the early adopters, let them talk to their neighbors, and watch the whole industry catch up quickly."

Technical Summary: Agent-Based Modeling of Low-Emission Fertilizer Adoption for Dairy Farm Decarbonisation

Problem Statement
Agricultural systems, particularly Irish dairy farming, are significant contributors to greenhouse gas (GHG) emissions, primarily driven by nitrogen fertilizer use, livestock management, and land-use practices. While practical low-emission measures exist—such as the adoption of protected urea (urea treated with the urease inhibitor NBPT)—adoption across farms is uneven. This heterogeneity stems from differences in management capacity, structural constraints, and social interactions. Traditional top-down assessment models often fail to capture this complexity, relying on assumptions of uniform behavior and immediate uptake while ignoring peer effects, path dependence, and stochastic decision-making. Consequently, there is a need for a framework capable of simulating the socio-technical evolution of decarbonization, treating adoption not merely as a financial calculation but as a dynamic process influenced by social contagion and structural farm characteristics.

Methodology
The study develops an empirically grounded Agent-Based Modeling (ABM) framework to simulate the diffusion of low-emission technologies across 295 Irish dairy farms over a 15-year horizon.

Data Foundation: The model utilizes a comprehensive farm-level dataset provided by Teagasc (Ireland's Agriculture and Food Development Authority), aggregated from the Irish Cattle Breeding Federation (ICBF) and Bord Bia. This dataset includes production performance (milk yield, solids, herd structure), input usage (fertilizer types, lime, feed), and economic variables.
Agent Structure: Each of the 295 farms is modeled as an autonomous agent with specific attributes: land area, livestock units, milk production, and nitrogen input. Initially, all agents use conventional Calcium Ammonium Nitrate (CAN).
Social Network Architecture: To replicate the social dynamics of agricultural discussion groups, the model employs a Watts-Strogatz small-world network. This structure features high local clustering (representing stable peer-advice circles) and short average path lengths (representing information transfer across groups via advisers). The network parameters ( $k=4$ connections, rewiring probability $p=0.1$ ) were fixed to represent a sparse, clustered peer structure rather than tuned to match specific adoption curves.
Adoption Mechanism: Adoption of Protected Urea is modeled as a stochastic, probabilistic decision process. The probability of adoption ( $P_{i,t}$ $P_{i, t}$ ) for a non-adopting farm $i$ $i$ at time $t$ $t$ is a function of:
1. Baseline inclination: A fixed propensity to innovate ( $\beta_0$ ).
2. Peer influence: A lagged signal ( $S_{i,t-1}$ ) representing the mean adoption rate of neighboring farms, weighted by a calibrated social influence parameter ( $\omega = 0.85$ ).
3. Structural factors: Normalized farm size and milk production.
4. Policy shocks: Exogenous adjustments for carbon taxes or subsidies.
Policy Scenarios: The model simulates three scenarios: a baseline (no policy), a carbon tax (€71/t CO2 applied to embedded fertilizer emissions), and a subsidy (€200/tonne on Protected Urea).
Emissions and Economics: The model calculates direct nitrous oxide emissions based on fertilizer-specific emission factors (CAN: 0.0149; Protected Urea: 0.0040) and converts them to CO2-equivalents. It also computes marginal abatement costs (MAC) by comparing private input cost changes against total abatement, distinguishing between private costs and social costs (including government expenditure).
Validation and Uncertainty: The model was calibrated using historical adoption data (2019–2024) and validated against out-of-sample data (2023–2024) using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and $R^2$ metrics. Robustness was assessed via 250 Monte Carlo iterations and sensitivity analyses on social influence weights and subsidy levels.

Key Contributions

Empirical Calibration: The study connects real-world farm-level data to a calibrated ABM, successfully modeling realistic diffusion pathways for fertilizer adoption with high predictive accuracy ( $R^2 = 0.979$ , RMSE = 0.0274).
Integration of Biophysical and Economic Modeling: It combines economic costing with biophysical emissions modeling to estimate marginal abatement costs within a simulated environment, distinguishing between private and social welfare costs.
Quantification of Uncertainty: By utilizing Monte Carlo methods, the study quantifies uncertainty in adoption timing and cumulative emissions, providing probabilistic insights rather than deterministic single-point estimates.
Policy Laboratory Framework: The model serves as a "policy laboratory" to pre-test the robustness and diffusion velocity of climate strategies, conceptualizing decarbonization as a socio-technical evolution rather than a purely monetary calculation.

Results

Predictive Accuracy: The model demonstrated high predictive accuracy, with an $R^2$ of 0.9793 and an RMSE of 0.0274. The inclusion of social network effects was critical; removing them increased the RMSE fivefold (from 0.0274 to 0.1476), confirming that peer-to-peer influence is a dominant mechanism in fertilizer adoption.
Adoption Dynamics: Adoption followed a logistic S-curve. The baseline scenario reached 50% adoption in ~7.2 years. The carbon tax accelerated this to ~3.6 years, while the subsidy scenario was the most rapid, reaching 50% in ~3.0 years.
Cumulative Abatement: Over the 15-year horizon, the subsidy scenario achieved 121.35 Gg CO2-eq avoided, approximately 24% higher than the carbon tax scenario (97.73 Gg CO2-eq). This difference is attributed to the earlier adoption peak under the subsidy, which prolongs the period of emissions savings.
Distributional Shifts: A Kolmogorov-Smirnov test confirmed that policy interventions fundamentally alter the sector's carbon intensity distribution ( $D = 0.2407, p < 0.001$ ), effectively suppressing the high-intensity tail and dragging "laggard" farms toward the benchmark mode.
Economic Viability: The private marginal abatement cost was calculated at €0.80 per tCO2-eq, significantly lower than current EU carbon price benchmarks. However, when including government subsidy expenditures, the social abatement cost rose to €4.54 per tCO2-eq.
Heterogeneity: Structural analysis revealed that carbon intensity is driven by nutrient-use efficiency and stocking balance rather than farm scale alone. "Laggard" farms (top 20% emitters) exhibited high stocking rates (~4.5 LU/ha) without corresponding milk yield gains, indicating diminishing returns and nitrogen loss.

Significance and Claims
The paper claims that its primary significance lies in providing a validated, generalizable framework for investigating heterogeneous, network-mediated technology diffusion in agricultural systems. By integrating empirical data with agent-based modeling, the study demonstrates that:

Social Contagion is Critical: Peer influence is not a secondary factor but a primary driver of adoption velocity. Policies that fail to account for network dynamics may underestimate the speed of diffusion or the impact of early adopters.
Subsidies vs. Taxes: While both carbon taxes and subsidies can drive adoption, subsidy-based interventions create faster diffusion and greater cumulative abatement due to their ability to trigger self-reinforcing social tipping points earlier in the transition.
Path Dependence: The model illustrates that policy timing interacts with network contagion to create durable transition pathways. Early incentives can "lock in" behavior by reducing perceived risks through peer validation, a dynamic that static price mechanisms (like taxes) may struggle to replicate unless they are sufficiently strong to cross the same social threshold.
Targeted Interventions: The framework suggests that targeting structurally influential, high-emission farms yields the greatest marginal impact, supporting the design of differentiated, evidence-based interventions rather than one-size-fits-all approaches.

The authors position the model as a tool for "policy laboratory" experimentation, enabling stakeholders to anticipate and optimize decarbonization strategies before implementation, while acknowledging limitations regarding bounded behavioral complexity and static economic variables.