HAICON26: AI as a Connecting Link for Environmental Research

The sixth Helmholtz AI Conference 2026 (HAICON26) brought together AI representatives from academia and industry in Munich from June 8–11, 2026. Organised by Helmholtz AI, the conference was held under the guiding theme “AI for Science” and focused on how modern AI methods can accelerate scientific research and enable new discoveries. The workshop “AI in Environmental Research,” led by Martin Schultz and Florentine Weber, focused on the role of AI as a connecting element between previously separate fields of environmental and Earth system research. Particular attention was given to agentic AI systems, foundation models for environmental applications, and the challenges and opportunities associated with a more integrated Earth system perspective.

From the ESDE research group, Martin Schultz, Florentine Weber, and Sebastian Buschow took part in the conference.

From Fragmented Knowledge to an Integrated Earth System Perspective

The workshop brought together around 25 participants from various Helmholtz centers and external institutions, including researchers from KIT, DKRZ, AWI, the University of Bonn, Fraunhofer, as well as representatives from industry, including NVIDIA.

The represented research areas ranged from weather and climate prediction, air quality research, flood forecasting, remote sensing, sea-ice classification, and biodiversity analysis to acoustics, agricultural applications, and technical challenges. This diversity vividly illustrated the increasing fragmentation of the environmental sciences: different disciplines operate with their own data structures, models, terminology, and assumptions, making a holistic view of the Earth system more difficult.

The central question of the workshop was therefore:

Can AI become the connective tissue of Earth system research?

In their introduction, Martin Schultz and Florentine Weber outlined the current challenges:

• Highly fragmented scientific disciplines
• Separate data infrastructures and metadata standards
• Incompatible spatial and temporal scales
• Static, manually developed workflows
• Limited domain-specific perspectives

In contrast, the vision of future AI systems is one in which different knowledge domains can be connected. The discussion focused in particular on agentic AI systems that could independently gather information, orchestrate tools, identify uncertainties, generate hypotheses, and selectively initiate simulations. Such systems could, in the long term, enable an integrated view of the Earth system and support the transition from pure prediction toward knowledge-based decision support.

As examples of current developments, Martin Schultz presented several initiatives from the ESDE community, including the foundation model projects WeatherGenerator, RAINA and HClimRep for applications of weather forecasting and climate predictions and the open data infrastructures of MeteoCloud und TOAR. Additionally, he highlighted the CESOC (Center for Earth System Observation and Computational analysis) Machine Learning Working Group and the MLESM26-Workshop (24-26 August).

World Café I: Cross-Domain Challenges and Opportunities

The dominant challenges associated with stronger integration across different environmental research domains were identified, from a natural science perspective, as the heterogeneity of data types, spatial and temporal scales, and the diversity of physical variables. A particularly difficult issue is not merely the integration of data itself, but the consistent coupling of processes across system boundaries—such as between the atmosphere, land surface, ocean, and biosphere—especially when different modeling assumptions and levels of uncertainty must be reconciled. Compounding this challenge is the lack of shared ontologies and semantic standards, which currently severely limits automated, cross-domain model coupling.

A central scientific challenge concerns the distinction between correlation and causality. Participants discussed how data-driven AI methods can be more closely integrated with physical knowledge in order to identify not only statistical relationships but also robust and transferable cause-and-effect mechanisms. Closely related to this is the question of how uncertainties and potential model hallucinations can be systematically detected, quantified, and controlled within decision-support and simulation workflows—particularly in scenarios where models act as active scientific agents within iterative research processes.

Agentic research architectures were discussed as a particularly promising future direction. In such architectures, AI systems would not only analyze data but also actively explore hypothesis spaces, propose potential data dependencies, and iteratively test them against physical models. This would create a closed scientific loop encompassing hypothesis generation, simulation, validation, and refinement. In this context, participants also considered the possibility that agents could autonomously formulate research questions and derive experimental designs, for example to identify previously unknown couplings between environmental and Earth system processes.

Another innovative approach involves multi-agent systems with domain-specific specialisation. Rather than relying on a single monolithic model, different “expert agents” representing fields such as atmospheric physics, hydrology, remote sensing, or energy systems could interact with one another to negotiate competing hypotheses and generate coherent cross-domain interpretations. Such systems could also make knowledge gaps between disciplines explicitly visible and therefore more amenable to systematic investigation.

Finally, the development of agentic benchmarks was discussed as an important next step. These benchmarks could go beyond traditional performance metrics and evaluate the extent to which AI systems are capable of independently developing, optimising, and validating research tools against physical reference models. This would make it possible to measure not only model quality but also the degree of autonomy in scientific workflows, including the extent to which human intervention remains necessary.

World Café II: ML Methods for Earth System Science

The second discussion round focused on the methodological foundations of future AI systems for environmental research.

An initial discussion centered on the role of large general foundation models versus specialized models. Both approaches were considered relevant; however, attention was drawn to the costs associated with maintaining this parallelism. In particular, participants discussed whether marginal performance gains - on the order of only a few percentage points - justify the significantly higher energy demands of large-scale models.

Data scarcity was not regarded as the primary challenge. Instead, it was emphasised that large volumes of Earth observation data are already available, but are often not in a form that can be directly used for AI applications. Data curation and the provision of AI-ready datasets were therefore identified as key challenges.

Additional methodological focal points included:

• explainable AI
• uncertainty quantification
• the role of standardized benchmarks in tension with domain-specific tasks
• digital twins and synthetic data
• and the question of how AI models behave in operational deployment

In the area of agentic AI, both issues of reliability and possible system architectures were discussed. A key open question was whether future solutions should rely on single high-performance models or on multi-agent systems with specialized components.

Furthermore, the role of large language models as scientific tools was addressed. They can accelerate research processes, lower the barrier to methodological access, and enable researchers without a strong machine learning background to conduct more complex analyses. At the same time, it was noted that a lower entry barrier does not automatically lead to better scientific processes, and there is a risk that users may lose focus on the core research question.

World Café III: Ethics & Equity – From Prediction to Decision Support

Who controls and understands AI systems when they evolve from predictive tools into autonomous decision-making partners? The third discussion round addressed the ethical, societal, and organizational aspects of future AI applications in environmental research, as well as their transition from purely predictive systems to decision-support tools.

A deep topic was the question of broad and equitable access to modern AI technologies. While open data, open-source models, and increasingly available computing resources are making environmental modeling more accessible, the development and training of large environmental foundation models remain concentrated in a small number of institutions with the necessary infrastructure. At the same time, participants expressed the expectation that open software ecosystems and freely available models could gradually break up this concentration. It was also discussed that large language models may in the future compensate for missing expertise and enable smaller research groups to access methods that were previously reserved for specialized centers.

Another key focus was the question of what forms of control and governance will be required for future AI and agent systems. Participants emphasized the need for transparent information about training data, model assumptions, operational constraints, and potential biases. At the same time, it was noted that modern agentic systems are becoming increasingly powerful while their internal decision-making processes are becoming harder to interpret. This creates the risk that users may base decisions on results whose derivation they only partially understand. The challenge was illustrated by the analogy of managing a team whose members are highly competent but largely opaque in their thinking and working processes.

Particular attention was given to the issue of biases and hidden influencing factors. In addition to well-known data biases, potential implicit constraints introduced by commercial providers were discussed, which may be embedded in models, training data, or development tools. As countermeasures, open models, open datasets, and scientific competition were highlighted, since only independent reproducibility enables systematic comparison of different approaches. It was further pointed out that many datasets exhibit geographic and cultural imbalances, with regions outside heavily monitored industrialized countries often being underrepresented.

In the context of bias, the importance of rare events was also emphasized. Extreme events are, by definition, outside typical data distributions and therefore pose a particular challenge for data-driven systems. The discussion made clear that AI systems should not be evaluated solely based on average performance metrics, but must also be assessed in terms of robustness and reliability in rare yet societally highly relevant situations.

Closely related to this was the question of how to deal with uncertainty. Participants advocated for a stronger culture of error awareness and uncertainty awareness, in which not only predictions but also their limitations are made visible. Both technical methods for representing uncertainty and organizational approaches were discussed, such as the targeted involvement of individuals whose role is to systematically identify weaknesses, flawed assumptions, and edge cases. Supportive user communities and open exchange platforms were mentioned as potential tools for making recurring issues and model failures visible at an early stage.

As an overarching conclusion, the demand for openness ran through the entire discussion. Open data, open models, and transparent development processes were not seen merely as a technical preference, but as a fundamental prerequisite for interpretability, trust, participation, and the long-term societal acceptance of AI-based environmental information and decision-support systems

Conclusion

The “AI in Environmental Research” workshop impressively demonstrated that environmental sciences are undergoing a fundamental transformation. While environmental research today is often organized in separate disciplines, the rapid development of foundation models and agentic AI systems opens up new possibilities for an integrated view of the Earth system.

At the same time, it became clear that the greatest challenges lie less in the models themselves than in data standards, interoperability, uncertainty assessment, and scientific validation. The discussions made it clear that AI should in the future not only be understood as a tool for prediction, but also as a potential mediator between scientific disciplines.

A further highlight at HAICON26 was Martin Schultz’s invited talk, “Rethinking Weather and Climate Modelling in the Age of AI,” which addressed the role of modern AI methods for the next generation of weather and climate modeling and placed the workshop’s themes into the broader context of Earth system research.