Integrated Assessment Models (IAMs)

• Integrated Assessment Models (IAMs) are computational frameworks that quantitatively couple climate, economic, and technological dynamics to assess environmental impacts and inform policy.
• They integrate modular components—including economic growth, emissions, climate response, and damage estimation—using approaches from neoclassical optimization to agent-based simulations.
• IAMs address uncertainty through stochastic simulation, robust optimization, and surrogate modeling, thereby refining metrics like the social cost of carbon for strategic planning.

Integrated Assessment Models (IAMs) are computational frameworks that quantitatively link key processes in the physical climate system, the biosphere, and socioeconomic systems to assess the impacts, mitigation, and adaptation pathways for global environmental change. They are fundamental for tracing causal chains from anthropogenic activities—such as energy production, land use, and industrial output—through greenhouse gas emissions and physical climate response, to sectoral and aggregate economic damages, and back to policy levers that modulate emissions or adapt to impacts. The IAM paradigm encompasses a spectrum from optimizing neoclassical equilibrium models (e.g. DICE, REMIND, MESSAGE-GLOBIOM), to detailed bottom-up simulation frameworks with heterogeneity and bounded rationality (e.g. E3ME-FTT-GENIE), to agent-based or reinforcement learning-driven negotiation simulators. IAM outputs are central to policy processes (including IPCC assessments) for quantifying the social cost of carbon (SCC), exploring integrated policy scenarios, and stress-testing the consequences of uncertainty or deep structural change.

1. Core Structure and Mathematical Formulation

Canonical IAMs decompose the coupled climate–economy–technology system into modular submodels:

• Economic Module: Typically a global or regional neoclassical growth model with Cobb–Douglas or CES aggregate production ($Y_t = A_t K_t^{\alpha} L_t^{1-\alpha}$), capital accumulation with depreciation, consumption–investment choices, and social welfare optimization via utility aggregation (e.g., $\max \sum_t \beta^t u(C_t)$) (Cai, 1 Nov 2025, Cai, 2020).
• Emissions and Technology Module: Emissions are generated as a function of economic activity and a structural carbon intensity parameter. Energy sector detail varies: bottom-up modules represent fuel-technology mixes subject to investment, learning, and resource constraints (e.g., MESSAGE, REMIND); top-down frameworks impose exogenous carbon intensity decay (Srikrishnan et al., 2019, Mercure et al., 2017).
• Carbon Cycle and Climate Module: Atmospheric CO$_2$ and other GHGs evolve in multi-reservoir box models ($M_{t+1} = A_M M_t + \text{emissions}$), coupled to simple global or two-layer thermal models ($$T_{t+1} = T_t + \xi_1 F(M^\mathrm{at}_t) - \xi_2 T_t$$), mapping radiative forcing to global mean temperature (Cai, 1 Nov 2025, Murakami et al., 16 Apr 2025).
• Damage and Abatement Cost Module: Economic damages are implemented via quadratic (or higher-order) functions of temperature ($D(T) = \alpha_1 T + \alpha_2 T^2$), acting as either level or growth-rate shocks. Abatement costs are commonly polynomial in the mitigation rate $\mu$ ($\text{Cost} = \theta_1 \mu^{\theta_2} Y$), and are subtracted from gross output (Biswas et al., 2 May 2025, Dai, 2023).
• Policy Interface: IAMs typically allow carbon taxes, cap-and-trade, technology subsidies, regulatory standards, or club/tariff structures to be simulated as exogenous levers or as endogenous controls under social (or decentralized) optimization (Dai, 2023, Biswas et al., 2 May 2025).

This core architecture can be further enriched with regional disaggregation, sectoral detail, and dynamic calibration to empirical data or process models (Murakami et al., 16 Apr 2025, Mercure et al., 2017).

2. Treatment of Uncertainty and Risk

Uncertainty is a structural and quantifiable feature of IAMs, manifesting in five principal categories:

1. Parametric Uncertainty: Unknown parameters (climate sensitivity, damage coefficients, behavioral elasticities) are represented by probability distributions; Monte Carlo ensembles and Bayesian calibration are used extensively (Cai, 1 Nov 2025, Srikrishnan et al., 2019).
2. Scenario Uncertainty: IAMs enable counterfactual exploration of exogenous trajectories (SSPs, technology pathways, land use, policy assumptions), each supplying alternative constraints or drivers (Cai, 1 Nov 2025, Murakami et al., 16 Apr 2025).
3. Structural (Model) Uncertainty: Competing module specifications or alternative functional forms (e.g. quadratic vs. higher-order damages, CES vs. physics-based energy transition) are bracketed or averaged (Sgouridis et al., 2016, Estrada et al., 2023).
4. Stochastic (Risk) Uncertainty: Explicit modeling of random shocks (e.g., productivity, climate tipping points) via recursive dynamic programming, stochastic Bellman equations, or approximate solution methods (Chebyshev collocation, time iteration) (Cai, 1 Nov 2025, Cai, 2020).
5. Deep (Ambiguity) Uncertainty: Robust control and max–min (worst case) and min–max regret criteria are applied when probabilities for key outcomes are contested or unavailable (Cai, 1 Nov 2025).

Methodologically, IAMs employ stochastic dynamic programming, Monte Carlo simulation, robust optimization, and surrogate modeling (e.g., Gaussian processes, manifold learning) to quantify and propagate uncertainty through output distributions (e.g., SCC, temperature, damages) (Cai, 1 Nov 2025, Zhang et al., 2020).

3. Advancements in Model Representations and Coupling

Recent developments expand standard IAM representations in several directions:

• Agent-Based and Reinforcement Learning Architectures: AB-IAMs replace the representative-agent equilibrium with heterogeneous, bounded-rational actors (households, firms, banks, governments), whose economy–energy–climate interactions and response to climate policy are governed by micro-founded decision rules and stochastic updating (Naumann-Woleske, 2023, Dai, 2023). Multi-agent RL formulations treat climate negotiation or emissions control as a cooperative-competitive game among autonomous regions, learning Pareto-optimal power-law policies via gradient-based RL (Biswas et al., 2 May 2025).
• Soft Coupling with Sectoral Models: System-level IAMs (e.g., REMIND) are bidirectionally soft-coupled with sectoral, high-resolution models such as power market dispatch (DIETER), with price and generation feedback ensuring cross-scale consistency between aggregate and hourly operation (Gong et al., 2022).
• Emulators and Machine Learning Surrogates: To enable rapid scenario projection beyond computationally expensive model runs, time-invariant marginal abatement cost (MAC) curves are fit to IAM outputs and coupled to reduced-complexity climate models, emulating GHG price-quantity behavior for extension to 2150 and beyond (Xiong et al., 2022, Xiong et al., 4 Dec 2025). Manifold sampling captures low-dimensional latent structures in high-D scenario spaces, supporting efficient risk quantification and decision analysis (Zhang et al., 2020).
• Policy Portfolio and Equity Assessment: New frameworks allow simultaneous multi-objective optimization (equity, GDP, temperature) and explicit representation of national or regional actors, exposing trade-offs between welfare and climate stabilization across the feasible policy Pareto front (Biswas et al., 2 May 2025).

4. Key Results, Applications, and Scenario Design

IAM analyses underpin assessments of future emissions trajectories, damages, policies, and their effect on climate stabilization:

• Social Cost of Greenhouse Gases: The canonical output is the SCC, representing the discounted marginal social damage of a ton of CO$_2$ (or CH$_4$, N$\max \sum_t \beta^t u(C_t)$0O). Including amplifying feedbacks (e.g., for methane: wetland emissions, OH scavenging) increases central SCC estimates by up to 44%, highlighting the need to update IAMs with Earth System Model-calibrated feedbacks (Colbert et al., 2020).
• Scenario Analysis and Policy Timelines: IAMs project that even under net-zero CO$\max \sum_t \beta^t u(C_t)$1 strategies by 2050, median warming exceeds 2°C; SCC values rise from $\max \sum_t \beta^t u(C_t)$2250–400/t by 2100, robust across six leading process-based IAMs (Murakami et al., 16 Apr 2025). The marginal SCC is relatively insensitive to the exact net-zero year, while peak warming is strongly affected.
• Climate Damages and Variability: Accounting for spatial (Svar) and temporal (Tvar) temperature variability increases projected damages and SCC by 25–30%, correcting low biases in models that absorb all warming into a global mean (Estrada et al., 2023). Explicit heterogeneity in climate damages and adaptation is thus critical for proper policy evaluation.
• Technology Diffusion Dynamics: Behaviorally realistic, simulation-based IAMs implementing evolutionary diffusion (S-curve) and multitarget policy packages (taxes, standards, subsidies, procurement) yield much more rapid low-carbon technology uptake (EVs, renewables) than optimization IAMs assuming a uniform carbon price (Mercure et al., 2017, Mercure et al., 2017).
• Long-Horizon Extensions: Computationally efficient emulators allow consistent analytic and scenario design to 2150, capturing the centennial-scale consequences of current investments under alternative policy extensions (linear price, zero-emissions, cumulative budget, or temperature-optimal) (Xiong et al., 4 Dec 2025, Xiong et al., 2022).

5. Critiques, Limitations, and Methodological Innovations

Several key critiques and challenges shape both the design and interpretation of IAM results:

• Energy Substitution Artifacts: Static CES functions for energy substitution overstress the difficulty of fossil phaseout, producing artifactual perpetual carbon price inflation and exponential abatement costs inconsistent with historical S-curve transitions. Dynamically varying elasticity or physics-based stock-flow models generate more realistic, non-monotonic cost paths and policy signals (Sgouridis et al., 2016).
• Equilibrium and Optimizing Limitations: Traditional neoclassical IAMs rely on optimizer agents with full information and perfect foresight. Simulation-based and agent-based models relax this assumption, allowing for bounded rationality, endogenous financial dynamics, technological lock-in, and emergent crises, better matching observed transition phenomena (Naumann-Woleske, 2023, Mercure et al., 2017).
• Uncertainty Handling and Policy Robustness: IAM projections and optimal policies are highly sensitive to deep structural uncertainties, such as non-linear climate feedbacks, damage functional forms, technological shocks, and tipping risks. Robust control, adaptive policy, and ensemble analysis are required to account for these deep uncertainties and design resilient strategies (Cai, 1 Nov 2025, Cai, 2020).
• Computational Scalability and Emulation: High-dimensional, stochastic-IAMs with regional, sectoral, and temporal granularity require advanced numerical methods (Chebyshev, sparse grids, surrogate models) for tractable solution and sensitivity analysis (Cai, 1 Nov 2025, Xiong et al., 2022).

6. Future Directions and Research Frontiers

Ongoing evolution in IAM methodology reflects new scientific and policy demands:

• Integration of High-Fidelity Process Modules: Seamless coupling to global vegetation, hydrology, land use, biodiversity, and sectoral impact models expands the physical realism and resolution (spatial, temporal, process) of IAM projections (Murakami et al., 16 Apr 2025, Gong et al., 2022).
• Extension to Multi-Century Horizons: Emulator-driven scenario extension to 2150 or beyond supports analysis of slow feedbacks (ice sheets, permafrost), and the robustness of transition pathways to long-run risks or shifting baseline conditions (Xiong et al., 4 Dec 2025).
• Exploration of Policy Mixes and Distributional Outcomes: New agent-based and reinforcement learning IAMs simulate the effect of interacting policy instruments, club negotiations, dynamic coalitions, and distributive impacts, supporting richer policy design and equity analysis (Biswas et al., 2 May 2025, Dai, 2023).
• Integration of Learning, Adaptation, and Endogenous Innovation: The next generation of IAMs seeks to endogenize learning-by-doing, R&D investment, adaptation choice, and policy learning, supporting adaptive strategies robust to emerging information and unanticipated crises (Mercure et al., 2017, Naumann-Woleske, 2023).

In aggregate, IAMs form an evolving, multi-paradigm ecosystem, central to quantitative climate-risk assessment and informing global policy on mitigation and adaptation under pervasive uncertainty. The discipline continues to synthesize advances from economics, climate science, data science, and computational modeling to provide increasingly robust, policy-relevant assessments of climate–economy futures.