1. Introduction

Climate change presents a pervasive global challenge with far-reaching implications for the entire planet. The theme of World Meteorological Day 2024, “At the Frontline of Climate Action,” emphasizes that catastrophic effects are imminent unless we act now. Addressing the challenges posed by climate change effectively hinges on several critical pathways. Foremost among these is the enhancement of our understanding of the trends in extreme weather events and their associated risks. This deeper insight is pivotal for developing precise and reliable predictions that are essential for proactive action and strategic planning. Additionally, leveraging advanced tools such as artificial intelligence (AI) stands out as a transformative approach. AI not only improves the accuracy of climate and weather predictions but also enhances our ability to model complex climate systems and simulate mitigation strategies. These methods collectively form a robust framework for tackling climate change, optimizing resource management, and ensuring communities can adapt to and mitigate the adverse effects of a changing climate.

Aligned with its core objectives, the Hong Kong University of Science and Technology (HKUST) hosted the “Climate, Weather, and Water Forum (CWWF) 2024” from 3 to 5 June 2024. This event expanded upon themes from the previous years—climate change, extreme weather, and water resources sustainability—to explore comprehensive strategies that address the interlinked challenges of climate, weather, and water, ensuring holistic environmental stability for future generations. Additionally, this year’s discussions extended to cover CO₂ emission reductions, carbon neutrality, and renewable energy in response to growing climate threats. A special session titled “AI4Climate” highlighted the integration of AI in these areas, showcasing its transformative potential for improving climate and weather forecasting, enhancing mitigation strategies, and developing innovative solutions for sustainable water management (Fig. 1).

View Full Size

Schematic overview of the forum’s main contents.

Citation: Bulletin of the American Meteorological Society 105, 11; 10.1175/BAMS-D-24-0240.1

• Download Figure
• Download figure as PowerPoint slide

The forum was attended in person by approximately 70 participants, with an expansive program comprising a total of 40 presentations—nearly double the count from the previous year. This comprehensive lineup encompassed 22 invited talks, each spanning 45 min, alongside 17 contributed talks, allotted 15 min each, and 12 poster presentations. The invited speakers are top scientists in the field of Earth sciences from leading universities and research institutions in the Americas, Europe, and Asia. Central to one of the forum’s overarching missions was the dissemination of cutting-edge scientific insights to governmental stakeholders. Noteworthy engagement was observed from several key agencies within the Hong Kong government, including the Hong Kong Observatory (HKO), the Civil Engineering and Development Department, the Water Supplies Department, the Environmental Protection Department, the Drainage Services Department, the Development Bureau, and the Environment and Ecology Bureau. These pivotal entities not only lent their support as organizational partners but actively participated in the forum’s deliberations. Notably, the head of HKO was invited to deliver the opening address, which encompassed discussions on pertinent topics such as Tropical Cyclone Maliksi, the unprecedented rainfall occurrences in southern China during the latter half of 2024, and the evolving applications of AI within the HKO’s operational framework. Additionally, the forum extended its reach to the Greater Bay Area, involving the Pearl River Water Resources Research Institute, further integrating regional expertise.

Throughout the 3-day forum, focused efforts were dedicated to achieving four primary objectives, each designed to explore different aspects of the overarching theme. On the first day, participants delved into enhancing our understanding of water resources and their sustainability. The second day shifted focus to the exploration of new technological opportunities in carbon neutrality and renewable energy. The third day was bifurcated into two sessions: the morning session concentrated on discussing extreme weather events and assessing the impacts of climate change, while the afternoon session was devoted to debating the latest advancements in AI technologies and their applications in climate science. Each morning session featured presentations from five distinguished experts with a panel discussion. These interactive sessions were designed to stimulate discussions on the salient themes expounded upon in the talks. On the afternoons of the first and second days, we organized two special sessions focused on discussing China’s national flood control strategies and policies, the development of renewable energy, and the implementation plans for carbon neutrality, following engaging the contributed talks and poster sessions. The posters, acting as visual aids, were prominently displayed throughout the forum, enriching the learning experience for all attendees, particularly early career fellows and students.

Both the structured arrangement from the forum organization committee and the active participation and contributions from forum participants ensured a comprehensive discussion of each critical area, shedding light on pivotal issues surrounding carbon neutrality, renewable energy, water resource management, weather and climate forecasting, and the evolving landscape of AI applications. In the subsequent sections, the contents covered during the forum will be concisely summarized in three distinct parts. The agenda of this forum is available online (https://cemlu.people.ust.hk/CWWF2024.html).

2. Carbon neutrality and renewable energy

Mitigating climate change effects and reducing greenhouse gas emissions accelerate the pursuit of carbon neutrality and the expansion of renewable energy sources such as wind power and photovoltaics. The increased focus on sustainability has led to a notable rise in interdisciplinary research aimed at achieving these goals, featuring a significant uptick in studies exploring innovative AI technologies.

Meteorologists play a crucial role in this field by merging insights from atmospheric science to optimize the design and operation of renewable energy systems. Sophisticated models like the Weather Research and Forecasting (WRF) Model and computational fluid dynamics tools such as Open-source Field Operations and Manipulations (OpenFOAM) are widely applied to forecast weather patterns and to assess their impact on energy production. Moreover, the integration of meteorological data with advanced modeling tools generates extensive big data, foundational for developing AI models that enhance energy yield optimization and grid resilience.

In the domain of solar radiation research, there has been a clear evolution through three developmental phases since the early 2000s. Initially, researchers focused on estimating solar radiation using data from ground stations, leading to the development of models based on spectral analysis (Yang et al. 2001, 2006, 2010). Technological advancements then shifted the focus toward satellite remote sensing methods (Tang et al. 2019a), which led to the creation of a comprehensive global solar radiation dataset with detailed spatial and temporal resolution, using International Satellite Cloud Climatology Project (ISCCP) cloud data (Tang et al. 2019b). This dataset has provided invaluable insights into the global distribution of solar radiation.

More recently, research has moved toward integrating solar radiation data using AI algorithms. For instance, a convolutional neural network–based homogenization technique has been applied to the ERA5 reanalysis dataset to homogenize a long-term satellite-derived solar radiation dataset (i.e., ISCCP-Institute of Tibetan Plateau Research) (Shao et al. 2022), enhancing the accuracy and reliability of solar radiation data. This improved data quality is crucial for enhancing the predictive accuracy and optimization of solar energy systems. These developments in meteorology for renewable energy and solar radiation research underscore the vital role of interdisciplinary collaboration in advancing the path toward carbon neutrality. By leveraging the expertise of atmospheric scientists, engineers, and data scientists, these fields are at the forefront of fostering innovation and refining renewable energy infrastructures. This collaborative approach is essential for addressing the complex challenges inherent in transitioning to a sustainable energy landscape and achieving global carbon neutrality goals.

3. Water resources and sustainability

Reservoirs have traditionally played a crucial role in managing streamflow to serve human and ecological needs. This role has become increasingly complex due to climate change and societal developments, leading to hydroclimatic uncertainties and concerns about water availability. These challenges, exemplified by groundwater depletion, deglaciation, and the reduction or disappearance of lakes, have led to higher demands on dwindling water reservoirs.

From a scientific perspective, the integration of AI technologies marks a significant advance in hydrological analysis and reservoir management (Hejazi and Cai 2011). Traditional machine learning methods, such as supervised learning, have been vital in extracting operational insights from historical reservoir data, highlighting the need for a versatile reservoir model that accurately reflects operational complexities while efficiently managing computational resources (Zhao and Cai 2020). A notable innovation is the hidden Markov decision tree (HM-DT) model, which combines hidden Markov models (HMMs) and decision trees. This model leverages HMMs to capture temporal dependencies and latent system states, while decision trees offer a clear, interpretable decision-making framework. By integrating these approaches, the HM-DT model effectively tailors operational strategies to specific reservoirs, considering their climate, functionality, and regulations. However, the opaque nature of these AI-driven models necessitates a shift toward more transparent and generic models (Chen et al. 2022).

AI’s role extends to uncovering latent patterns and enhancing understanding of hydrological processes, improving the performance of hybrid physical-AI models. For example, long short-term memory networks have been used to analyze spatiotemporal flood peak formation across the continental United States, revealing dynamics mainly driven by streamflow, independent of subjective factors like seasonality (Jiang et al. 2022). AI models often outperform traditional process-based models (PBMs) in various tasks, though PBMs maintain advantages in physical coherence and generalizability. To bridge this gap, a specialized recurrent neural network has been developed that incorporates numerical solutions from PBMs, demonstrating the potential of physics-informed deep learning models for accurate and transferable predictions (Jiang et al. 2020).

In engineering, particularly in China, intensified human activities have led to a “dual natural–social water cycle.” Urbanization has transformed agricultural lands and water bodies into impermeable surfaces, reducing infiltration and increasing surface runoff. This transformation accelerates floodwater flow, intensifying flood events. Addressing these challenges involves a comprehensive strategy that includes using upstream reservoirs to retain floodwaters, employing urban dikes and embankments to divert floodwaters, maintaining and enhancing urban drainage systems, implementing sponge city concepts to improve water absorption, and enhancing disaster response mechanisms to mitigate flooding impacts effectively.

4. Climate, weather, and extremes

The realms of climate prediction and weather forecasting persist as pivotal focal points of interest. The overarching objective of climate prediction lies in prognosticating forthcoming climatic patterns based on evolving boundary conditions, notably including factors such as greenhouse gas emissions. A comprehensive comprehension of the dynamic processes at play forms the bedrock for the development of predictive models. Notably, El Niño–Southern Oscillation (ENSO) stands out as a primary wellspring of predictability within the global climate system, with its characteristics and teleconnections with the broader climate canvas exhibiting continual evolution over the past century. An intriguing observation pertains to the occurrence of multiyear La Niñas, which wield considerably prolonged impacts in comparison to single-year occurrences, with a tally of 10 instances documented in the last century, notably five of which unfolded in the most recent quarter century (Wang et al. 2023). One prevailing conjecture posits that alterations in the Pacific mean state, whether instigated by external forcings or internal variabilities, can modulate the occurrences of El Niño and La Niña phenomena (Wang et al. 2023). Given the intricacies of the underlying physical mechanisms and the ongoing dynamics characterizing ENSO, the predictability of this phenomenon looms as a persistent and formidable challenge in the foreseeable future. Currently, two notable recent research breakthroughs have emerged about ENSO predictability. Zhao et al. (2024) have propounded an extended nonlinear recharge oscillator (XRO) model, demonstrating the capacity to furnish precise ENSO forecasts up to 16–18 months in advance. Additionally, Ham et al. (2019) have introduced a convolutional neural network, capable of generating accurate ENSO predictions up to 18 months ahead. These advancements stand as a testament to the ongoing endeavors within the scientific community to enhance the predictive capabilities surrounding ENSO dynamics, underscoring the interdisciplinary nature of climate prediction research and the pivotal role of advanced methodologies in advancing the frontiers of climatic forecasting.

In terms of weather forecasting, AI models have emerged as notable assets, showcasing remarkable prowess in this domain. The unveiling of the PanGu model in November 2022 (Bi et al. 2023) marked a significant milestone, underscoring the potential inherent in AI-driven prediction models. Subsequently, a succession of meteorological models have been introduced, each contributing to the advancement of forecasting capabilities. A pivotal development occurred in December 2022 with the inception of GraphCast (Lam et al. 2023), representing the first utilization of graph neural networks within meteorological models and extending forecast horizons to 9.75 days. The unveiling of the FengWu model in April 2023 (K. Chen et al. 2023) further elevated the standards, surpassing GraphCast by accurately predicting 80% of the 880 identified predictands and pushing the boundaries of skillful global medium-range weather forecasts beyond a 10-day lead time. The advent of 2024 witnessed the introduction of FengWu-Global High Resolution (GHR) as the pioneering AI-based global medium-range weather forecasting model operating at a resolution of 0.09°. In June 2024, the release of FuXi-2.0 extended the forecast lead time to an impressive 15–42 days. Noteworthy instances include the adoption of a two-step U-Net for forecasting summer precipitation and 2-m air temperature across the mid-lower Yangtze River region, the utilization of long short-term memory (LSTM) for Arctic sea ice predictions (Wei et al. 2022), the application of convolutional LSTM (ConvLSTM) for Atlantic Ocean wave forecasts (Ouyang et al. 2023), and the deployment of cycle-consistent generative adversarial network (CycleGAN) to refine summer precipitation forecasts over China (Yang et al. 2023, manuscript submitted to Geophys. Res. Lett.). While the training of large-scale AI models may demand substantial computational resources, once trained, these models typically exhibit accelerated computational speeds and enhanced user-friendliness in comparison to traditional physical-based models. Notably, forecast products encompassing variables such as wind fields and precipitation, among others of the upcoming 15 days from both the PanGu and FuXi models, are now accessible via the “Earth Weather” website maintained by the Hong Kong Observatory (https://maps.weather.gov.hk/wxviewer/index.html?lang=en).

The domain of subseasonal to seasonal (S2S) forecasting, serving as a crucial link between short-term weather predictions and long-term climate projections, emerged prominently as a leading-edge challenge in numerical modeling in the previous year (Zhang et al. 2023). The provision of real-time calibrated probabilistic subseasonal rainfall and temperature forecasts, facilitated by the subseasonal experiment (SubX) models developed by the International Research Institute, stands out for its capability to furnish insights into pivotal climate attributes such as the timing of seasonal rainfall onset crucial for agricultural planning, as well as the identification of risks associated with extreme precipitation events or heatwaves, thereby aiding public health preparedness. This framework not only opens avenues for the creation of tailored climate services but also fosters collaborative engagements with diverse user communities. The efficacy of S2S forecasts is also subject to episodic variability contingent upon the diverse drivers influencing the weather system. A notable advancement in this domain is the FuXi-S2S model, which enhances the precision of predicting the Madden–Julian oscillation (MJO) beyond the 30-day horizon achieved by the ECMWF S2S model, extending it to a commendable 36 days. This pioneering AI-driven FuXi-S2S model marks a watershed in medium-range and subseasonal forecasting (L. Chen et al. 2023).

5. Keywords of CWWF 2024: Prediction, AI, and disruptive work

Prediction is fundamental across all climate services, with a critical emphasis on bridging the gap in S2S time scales. AI, bolstered by the development of successful models in 2023–24, opens new frontiers and provides innovative means to enhance these predictive efforts in Earth sciences.

Enhancing prediction accuracy in capturing complex hydrometeorological events and improving S2S forecasting capabilities are pivotal areas ripe for advancement. However, challenges persist, such as ensuring the validity of AI-derived predictions based on historical data when applied to future forecasts in a dynamically changing climate. This issue is compounded in hydrological forecasting, where anthropogenic changes like urbanization and infrastructure development continuously alter watershed characteristics, thereby impacting both scientific research and practical engineering applications. A central concern is whether AI models trained on past data can maintain their robustness in predicting future conditions.

The role of AI in this predictive landscape is increasingly critical. The selection of the appropriate AI model tailored to address specific challenges is crucial. For instance, the integration of domain-specific generative pretrained transformer (GPT) models designed for climate-related applications illustrates significant progress. The evolving perception and capabilities of these AI models, like the latest ChatGPT-4o, demonstrate their potential in handling diverse data types—from textual to spatiotemporal information—thus laying a robust foundation for addressing a wide spectrum of atmospheric science challenges (Zhang et al. 2024). Despite their high predictive accuracy, the challenge remains to translate these forecasts into actionable insights, especially in areas crucial for governmental risk management.

The confluence of physical models with AI frameworks represents a promising avenue for further investigation and development. Beyond shedding light on latent physical processes (Jiang et al. 2020) and elucidating intricate mechanisms (Li et al. 2023), a critical inquiry emerges: how can AI models contribute to the refinement and augmentation of existing physical models? Moreover, the persistent challenge of limited observational data poses a significant impediment to the progress of hydrometeorology. Despite the application of advanced AI methodologies for data analysis, the absence of robust observational data undermines the efficacy of these techniques in delivering meaningful insights. The participants of the forum also reached a consensus that AI still has more potential in weather to S2S prediction, the limitation of data hinders its advancement in long-term climate forecast.

Disruptive work in this field extends beyond merely advancing AI technologies; it necessitates a fundamental reevaluation of our scientific inquiries, hypotheses, and methodologies to ensure they adequately address the complexities of contemporary challenges. Given the propensity of AI systems to “hallucinate”—a phenomenon often triggered by flawed training data, inadequate model architectures, or suboptimal training regimes—it is crucial for researchers to approach their work with increased scrutiny and critical thinking.

In harnessing the formidable capabilities of AI models, it is imperative that we not only pose more discerning questions but also deepen our understanding of AI mechanisms, regardless of our primary field of expertise. This deepened understanding also fosters cross-disciplinary collaboration and learning, aligning with the objectives of this forum and shaping the themes of our 3-day event, which has promoted engaging cross-disciplinary dialogue among participants. By broadening our knowledge base, we can more effectively identify potential flaws in model designs and training methods. Moreover, a comprehensive grasp of AI functionalities will allow us to more effectively articulate why certain models succeed where others fail, ensuring that our scientific endeavors remain both innovative and grounded in reality.