3MIP — Mobility and Migration Modeling Intercomparison Project

Different models. Same question. One coordinated answer.

3MIP is a community of researchers who apply their climate-migration models to a shared benchmark case. The first case is coastal Bangladesh.

Why the same question yields different answers

Different model architectures — agent-based, gravity, radiation, integrated assessment, and machine learning — encode different assumptions about how people decide to move. Run the same Bangladesh migration question through each and you get different answers. 3MIP makes that divergence visible and synthesizes it.

Fig. 01

Six architectures, one question

Net internal migration, coastal Bangladesh, 2025–2050, SSP2-4.5

01 Agent-based

Agent-based architecture, same question. About twenty dots of varying size, denser along the southwest coast and the corridor toward Dhaka; each dot is one simulated household cluster. In motion, dots drift toward the corridor and a few relocate.

02 Gravity

Gravity architecture, same question. Flow arcs of varying thickness from five coastal origin points to three destination cities; thicker arcs are larger modeled flows. In motion, arcs draw in and pulse with flow weight.

03 Radiation

Radiation architecture, same question. Nested rings spreading from one focal point, spacing widening and color fading outward — distance-decaying catchments. In motion, rings ripple outward in sequence.

04 Integrated assessment

Integrated assessment architecture, same question. Five regional polygons, each filled with a single flat intensity value; no detail inside regions, because the model aggregates. In motion, fills jump between discrete time-step states.

05 Machine learning

Machine learning architecture, same question. Smooth nested contour bands deepening toward a coastal core — a continuous predicted surface with no hard edges. In motion, the surface crossfades between predicted states.

06 Cellular automaton

Cellular automaton architecture, same question. A coarse grid of discrete cells in stepped intensities, clipped to the country outline. In motion, cells change state in discrete generations.
Net migration intensity low high
≈ 100 km
SSP2-4.5
Illustrative. Synthetic geometry tuned to show the spatial grammar each architecture tends to produce; 3MIP's purpose is to make this divergence legible across real participating teams. Archetype lineages: Bell et al. 2019; Simini et al. 2012; Stouffer 1940; Rigaud et al. 2018; ISIMIP3 protocol; Schiavina et al. 2019.

Long-form text alternative. Each of the six panels above shows the same stylized outline of Bangladesh, but drawn in the visual grammar of a different model architecture. Agent-based: about twenty individual dots of varying size — household clusters — denser along the southwest coast and the corridor toward Dhaka. Gravity: flow arcs of varying thickness from five coastal origin points to three destination cities; thicker arcs are larger flows. Radiation: nested rings spreading from one focal point, spacing widening and color fading outward. Integrated assessment: five regional polygons each filled with a single flat value — no interior detail, because the model aggregates. Machine learning: smooth nested contour bands deepening toward a coastal core, with no hard edges. Cellular automaton: a coarse grid of discrete colored cells. Panels share a sequential color ramp from low to high intensity, a scale bar, and the common research question. When animated, each panel moves in its own model's logic across a repeating 2025–2050 loop: the dots drift toward the corridor and a few relocate; the arcs draw in and pulse with flow weight; the rings ripple outward in sequence; the region fills jump between discrete time-step states; the contour surface crossfades between predicted states; and the cells change in discrete generations. A pause control stops the loop; with reduced motion enabled the figure is fully static.

Case 1: Coastal Bangladesh

Bangladesh has 175.7 million people, much of it living in a low-lying delta where sea-level rise, salinity, and flood dynamics interact with economic and social drivers of migration. Bell et al. (2021) projected continued migration toward Bangladesh's coast through 2100 under all studied scenarios — not away from it. The empirical density and policy stakes make this the right place for a first benchmark.

Read the case in full .

Next milestone

iEMSs 2026 — University College Dublin, July 12–16, 2026

Session C7: Mobility and Migration Modeling Intercomparison Project (3MIP) – An open, first synthesis

Workshop WSC7: Panel Discussion – 3MIP

The first public synthesis of 3MIP results. Open to all conference attendees. Session day and room follow the preliminary program, which publishes June 15, 2026.

3MIP at iEMSs Dublin Session details on the iEMSs program

How 3MIP works

  1. Register

    Tell us your modeling approach and research interest.

  2. Receive curated data

    Get access to the harmonized Bangladesh input dataset.

  3. Run your model

    Apply your approach to the shared question. Your model stays yours.

  4. Contribute to synthesis

    Submit outputs to the Climatic Change topical collection and present at iEMSs.

Leadership

  • Andrew Reid Bell

    Schleifer Family Professor of Sustainability, Department of Global Development, Cornell University

    Co-lead

  • Kelsea Best

    Assistant Professor, Department of Civil, Environmental & Geodetic Engineering and Knowlton School (City & Regional Planning), The Ohio State University

    Co-lead

  • Lars Tierolf

    PhD researcher, Institute for Environmental Studies (IVM), Vrije Universiteit Amsterdam

    Co-lead

Project coordination — Mario Keputa (Cornell University), mk2674@cornell.edu

Register

Register to participate

Open to climate-migration modelers using ABM, gravity, radiation, IAM, or ML approaches — and to domain experts on Bangladesh and climate-related mobility. No fee. Participation expectations on the Participate page.

Inspired by AgMIP and ISIMIP. Founded at Princeton C-PREE, September 2024.