New study in Nature Cities maps machine learning research for urban climate change mitigation

Staying below the 1.5°C threshold requires massive urban decarbonisation: adopting clean energy, expanding public transport, curbing urban sprawl, and implementing nature-based solutions for cooling and resilience.

With urban areas as the dominant form of human settlement, many are asking: can machine learning (ML) help make cities carbon-neutral? And more importantly, is current research on ML for climate change mitigation actually aligned with the most impactful approaches?

A global map of ML and urban climate research

A new study – “A systematic map of machine learning for urban climate change mitigation” published in Nature Cities and authored by Marie Josefine Hintz (Technische Universität (TU) Berlin, Hertie School, Potsdam Institute for Climate Impact Research (PIK), Nikola Milojevic-Dupont (PIK), Felix Creutzig (PIK, TU Berlin), Tim Repke (PIK), and Lynn H. Kaack (Hertie School) – takes a first global look. Marie Josefine Hintz is a PhD student in the AI and Climate Technology policy group, led by Professor Lynn Kaack, who is affiliated with the Data Science Lab and the Centre for Sustainability. 

The authors analysed 2,300 peer-reviewed articles published between 1994 and 2024. Since 2012, the field has experienced explosive growth. Most studies focus on:

• Transport (41%) – shared bikes and scooters, micro-mobility, traffic routing, electric vehicles

• Buildings (25%) – energy consumption, retrofits, occupant behaviour

These sectors are also where city governments have the largest opportunity to cut GHG emissions.

Where ML research misses the mark

This research highlights a critical gap: machine learning research often doesn’t align with the areas that could deliver the greatest impact on climate change mitigation. Some topics, like traffic management and consumer behaviour, are heavily studied. Interestingly, their effect on reducing greenhouse gas emissions is relatively small. Meanwhile, areas with far greater potential – such as redesigning street networks or switching fuels in buildings – remain underexplored. These are the kinds of shifts that could make a real difference.

In other words, researchers working in this area often pursue data where it is available or commercially attractive, rather than where ML could deliver the greatest reductions in GHG emissions.

A map with gaps

Geographically, the research is also unevenly distributed, with 83% of studies focussing on East Asia, North America and Europe. Research on cities like Beijing, China (121 studies) and New York City, USA (117) dominates.

Meanwhile, Africa, South Asia and Latin America, regions where urban growth is fastest, remain underrepresented. This risks perpetuating digital coloniality: exporting solutions designed for data-rich, high-income megacities to contexts where they don’t fit or local communities do not want them.

From research to reality

To realise ML approaches for climate action in practice, more work is needed. Just because a solution is technically feasible does not guarantee it will resolve real-world challenges without adverse consequences. The authors emphasise the risk of potentially distracting financial and personnel resources from other non-ML-based climate action. Quantification of impact is largely lacking in existing research studies, making it difficult for practitioners to assess costs and benefits. 

A path forward

The authors make recommendations for future research. In short:

• Focus on high-impact but neglected areas (like street networks).

• Measure and report real GHG emission reductions, rather than just quantifying model accuracy.

• Align research with the actual needs of policymakers and urban planners.

• Address regional gaps in data coverage and analyse opportunities for ML in climate action in low-income, data-scarce contexts to prevent reproducing patterns of digital coloniality.

• Engage with cross-disciplinary networks (e.g., Climate Change AI (climatechange.ai), Urban AI (urbanai.fr), or AfriClimate AI (africlimate.ai)) and build the knowledge to identify high-impact approaches.

The message is clear: ML is not a silver bullet. However, when coupled with climate policies, measures to ensure equity, and real-world challenges, it can become a powerful tool to support the decarbonisation of cities. Research plays a vital role in guiding ML for climate applications toward where emissions reduction matters most. Otherwise, we risk widening the gap between where knowledge is produced and where it is most urgently needed.

Read the full research paper here. 

Photo by Steve Johnson on Unsplash.

The Hertie School is not responsible for any content linked or referred to from these pages.