AI for sustainability sciences – a literature review

Introduction

This chapter summarizes our AI-assisted literature overview, based on a combination of strategic searches in literature databases, expert assessments, and AI-supported analysis. The aims of this literature analysis are twofold. First, to generate a high-level overview of AI methods across the selected sustainability issues based on keyword frequency in abstracts. Second, to compile a set of relevant scientific literature to complement subsequent expert-authored sections of this report.

While far from perfect, this approach allows us to map methodological innovations, application trends, and knowledge gaps at the intersection of AI and sustainability science. The analysis builds on the methodological taxonomy introduced earlier in the report, with a special emphasis on democratization—defined as the process by which access to, understanding of, and application of AI methods expanded beyond specialized research communities into broader scientific practice.

Results

We distinguish three generations of AI methods in our classification:

• Classical machine learning (ML). Dominant before widespread GPU adoption for training (~pre-2010). Characterized by traditional statistical and algorithmic methods without deep neural networks, primarily trained on central processing units (CPUs). This era saw increased access through libraries like scikit-learn, enabling broad scientific use.
• Focused deep learning (DL). Defined by the democratization/growing public access of DL starting around 2010–2012, enabled by GPU advancements and marked by superhuman performance in specific, narrow tasks (e.g., DanNet 2011,¹ AlexNet 2012²). While core concepts (e.g., convolutional and recurrent neural networks [CNNs and RNNs, respectively], and backpropagation) predate this, their practical feasibility and widespread adoption surged in this period.

New generation of AI. Characterized by the democratization of generative capabilities and foundation models since ~2020 (e.g., ChatGPT, Stable Diffusion), making advanced AI accessible to non-experts via prompting. While generative AI (GANs, VAEs, Diffusion) and transformers are central, this era also encompasses critical advancements like multimodality (e.g., CLIP), knowledge-guided/physics-informed ML, explainable AI (XAI), and smart robotics.

The use of AI methods in sustainability research has been substantive for the five years analyzed. A total of 5,603 papers (out of the total 8,504 included in the literature review, see Methods) were successfully classified into the AI generations listed above. The distribution shows that classical ML is the dominant approach for all the sustainability issues selected (2,381 articles), followed by focused DL (2,163 articles), and new generation AI (1,059 articles) (Fig. 3).

There are, however, distinct methodological patterns that differ between areas. Articles belonging to the issue area Enhancing Nature’s Contributions to People, for example, show the frequent use of AI methods like CNNs. This could be a reflection of frequent uses of AI, for example for species identification tasks. Neural networks and classical ML methods such as random forests and support vector machines (SVMs) are also widely used, for example in ecological modeling contexts. Transformers appeared infrequently, suggesting a more limited adoption (Fig. 4e).

AI methods used in the issue area Collective Decisions for a Planet under Pressure feature neural networks and random forests, alongside notable occurrences of transformers, LLMs, and reinforcement learning (Fig. 4h). The issue area Understanding a Complex Earth System shows frequent uses of recurrent CNN and long shortterm memory (LSTM), alongside persistent SVM and regression methods (Fig. 4b). The area Preparing for a Future of Interconnected Shocks displayed high occurrences of CNNs and LSTMs, potentially linked to hazard prediction applications, followed by frequent mentions of generative adversarial networks (GANs) and SVMs (Fig. 4a).

Stewarding Our Blue Planet revealed frequent uses of neural network methods, alongside unsupervised techniques such as principal component analysis (PCA) and k-means clustering (Fig. 4c). Improving Sustainability Science Communication was characterized by natural language processing (NLP) terms, particularly transformers and LLMs (e.g., BERT), implying a strong focus on language-based methods (Fig. 4g).

Prospering on an Urban Planet displayed strong uses of CNN and GAN methods followed by frequent SVM and regression techniques (Fig. 4f). Securing Freshwater for All demonstrated high occurrences of LSTMs and random forests, potentially indicating their use in hydrological forecasting applications, with CNNs also frequently represented (Fig. 4d).

As a general pattern, older AI methods seem to be more commonly used in the areas of Understanding a Complex Earth System, Securing Freshwater for All, and Stewarding Our Blue Planet.

Newer AI methods can facilitate participation by non-specialist stakeholders and showed notably higher adoption in the domains of Prospering on an Urban Planet, Improving Sustainability Science Communication, and Collective Decisions for a Planet Under Pressure.

Method Summary

The scoping literature search used the OpenAlex³ database for peer-reviewed articles published between January 2020 and December 2024. Search queries combined AI-related keywords (e.g., “machine learning,” “deep learning,” “generative AI”) with sustainability science concepts across the eight predefined focus areas (e.g., Nature’s Contributions to People, Stewarding Our Blue Planet, Prospering on an Urban Planet) defined by researchers from the Stockholm Resilience Centre and Potsdam Institute for Climate Impact Research (full search strategy in Appendix 2).

After removing duplicates, 21,648 articles remained and were thus included in the first step of the analysis. We later employed a tiered screening approach and selected the 1,670 most-cited articles (citation count >10 for articles published between 2023 and 2024, and citation count >30 for articles published between 2020 and 2022). These underwent manual title/ abstract review against predefined inclusion/ exclusion criteria (Appendix 3). Each included article was categorized as belonging to the single most relevant sustainability science issue area. The remaining 19,978 articles were screened using Rayyan,⁴ an AI-assisted systematic review tool, which flagged 7,679 articles as “Likely” or “Very likely” to be relevant. This yielded a final corpus of 8,504 articles combining manually selected and AI-flagged papers.

For the AI method classification of each paper, we used a locally deployed large language model (LLM) (DeepSeek-R1-7B). The model was used to extract AI methodology keywords from article abstracts. Extraction was guided by a taxonomy of AI and representative examples (Appendix 4). The taxonomy included keywords mapped to consolidated categories to prevent duplication (Appendix 5). Our manual validation on 100 randomly sampled articles demonstrated 93% accuracy in keyword extraction, and 88% agreement in method categorization. No formal quality assessment of individual papers was performed, as the objective was to characterize broad methodological trends and application breadth rather than assess individual study quality. Note that all methodological details of this review can be found in the appendices.

Limitations

This literature analysis has several limitations. First, our use of Rayyan, a proprietary AI screening tool, meant we couldn’t access or verify its internal algorithms, training data, or decision thresholds. The limited access could potentially introduce selection bias for lower-citation articles. Second, all screening relied solely on titles and abstracts rather than full texts, risking oversight of relevant studies where key details (e.g., methodology or sustainability links) appear only in the main content. Third, while manual validation showed strong accuracy, the LLM-based keyword extraction depended heavily on abstract clarity; ambiguous phrasing or niche methods might have been missed or misclassified.

To address these limitations, we implemented several mitigation strategies throughout the review process. To reduce potential selection bias from Rayyan’s screening, we confirmed the relevance of selected studies through title keyword analysis, providing supporting evidence that the included articles aligned with the intended scope of the sustainability issue areas. Nonetheless, we acknowledge that some domains, particularly those with consistently low representation (e.g., Stewarding Our Blue Planet and Improving Sustainability Science Communication) may still be subject to undersampling.

To counter limitations of LLM-based keyword extraction, we manually validated a random sample of 100 articles, assessing accuracy in AI method categorization and finding high accuracy rates (93% accuracy in keyword extraction, and 88% agreement in AI-generated categorization of AI-methods). While reliance on abstracts and LLM-assisted extraction cannot fully replace exhaustive full-text review, these iterative validation steps served as robust checkpoints for both inclusion and classification.

Despite the limitations, we believe that this review offers valuable broad-scale insights into AI’s evolving role in sustainability science.