AI-driven decision-making for water resource planning and hazard mitigation using automated multi-agents Open Access

Water resource planning and hazard mitigation are vital for sustainable water management, aimed at addressing hydrological risks like floods, droughts, and water pollution. These hazards pose significant socio-economic and environmental threats, making comprehensive planning essential. While traditional hydrological models have provided mathematically optimal solutions, they often overlook social and public dimensions crucial for long-term success and acceptance (D'Alpaos & Bottacin 2021; Chang et al. 2023).

Traditional hydrological models face critical limitations in addressing the social dimensions of water management (Daniell et al. 2023). These approaches typically prioritize physical and economic parameters while minimizing community values and preferences, involve stakeholders only after technical solutions have been developed rather than throughout the planning process, and struggle to simulate the complex, multi-dimensional decision-making processes that characterize real-world community planning (Kamyab et al. 2023). This gap between technical optimization and social acceptance represents a critical barrier to effective water resource management.

Recent advancements in artificial intelligence (AI) have revolutionized water resource management by enabling the processing of large datasets and simulating complex systems in flood prediction and management through remote sensing (Hammam et al. 2020; Motta et al. 2021; Pennings 2024). AI-driven tools such as AutoGen (Wu et al. 2023), GPT-4o-mini (OpenAI 2024), and DarkIdol-Llama-3.1-8B (QuantFactory 2024) are being incorporated into decision-making frameworks (Kadiyala et al. 2024; Martin & White 2024; Samuel et al. 2024; Vald et al. 2024), offering more adaptive and responsive strategies. This integration helps decision-makers balance hydrological challenges with community needs and social concerns while contributing to a more holistic disaster mitigation strategy encompassing environmental risks like wildfires, flooding (Algarni 2023; Alabbad et al. 2024) and droughts (Ghaffarian et al. 2023; Elshaikh-Hayaty et al. 2024; Salimi et al. 2024).

Recent studies have demonstrated significant advances in AI applications for water management. Zhao et al. (2025) propose digital twins and IoT-driven monitoring for water reserves management, enhancing visualization and adaptation strategies through immersive technologies. Ugural et al. (2024) apply fuzzy logic to address uncertainties in AI adoption decisions, identifying organizational factors crucial for technology implementation success. However, these approaches primarily focus on technical advancement and have not adequately addressed the social-technical divide in water resource management decision-making.

Social hydrology emphasizes the integration of community input into decision-making processes, exemplified by the multi-hazard tournament (MHT) framework. This framework bridges the gap between technically sound and socially acceptable solutions by simulating community involvement in hazard mitigation decisions (Konar et al. 2019; Saravi et al. 2019; Yousefi et al. 2020). Technology facilitates knowledge dissemination and engagement in disaster mitigation efforts (Sermet & Demir 2018; Le & Nguyen 2022), as demonstrated by AI applications in professional contexts like the US Fundamentals of Engineering exam (Pursnani et al. 2023; Sajja et al. 2024a). This study introduces an AI-driven MHT framework that leverages advanced models to simulate decision-making processes in water resource planning and hazard mitigation (Salimi et al. 2024; Sharma et al. 2024; Vekaria & Sinha 2024). The framework uses AutoGen to manage multi-agent interactions, with each AI agent representing a community stakeholder with unique attributes and priorities. Large-scale models enable structured discussions, evaluations, and negotiations (Bowes et al. 2021a; Sajja et al. 2023a), while exploring how AI-driven serious gaming can improve hazard mitigation planning (Teague et al. 2022; Chang et al. 2023). The research evaluates the framework's capacity to increase stakeholder awareness, understanding, and collaboration in water resource management, demonstrating how these technologies can facilitate teamwork and address challenges in educational and problem-solving contexts (Sajja et al. 2023b, 2024b, c; Farzana et al. 2024).

By integrating advanced AI technologies with socio-hydrology, this study provides a robust framework for understanding and addressing the complex interplay between technical solutions and public preferences in water resource management. The dual focus on public preferences and community response simulation offers valuable insights into social dynamics influencing flood mitigation strategies, which can guide policymakers and planners in designing more effective and socially acceptable water management strategies (Bowes et al. 2021b; Farhaoui & El Allaoui 2024; Piemontese et al. 2024; Project Performance International 2024; Qin et al. 2024).

To address these identified gaps, this study introduces an AI-driven MHT framework that makes three key contributions: (1) developing a novel integration of AutoGen and DarkIdol-Llama-3.1-8B to create a multi-agent simulation environment capable of representing diverse community stakeholders with distinct demographic and behavioral characteristics; (2) implementing a comprehensive decision-making framework that captures both technical and social dimensions of flood mitigation planning; and (3) providing quantitative analysis of how demographic factors influence strategy selection and budget allocation in hazard mitigation planning.

This study employs the MHT framework to explore and analyze decision-making processes in water resource management and hazard mitigation. By leveraging advanced AI technologies, the framework simulates complex, real-world scenarios involving multiple stakeholders with diverse backgrounds and preferences. The methodology integrates both technical and social dimensions into a comprehensive decision-making process.

The MHT framework addresses limitations in traditional hydrological models by incorporating social and public dimensions crucial for successful implementation of mitigation strategies (Kamyab et al. 2023). While these traditional models provide mathematically optimal solutions, they often overlook community preferences and social dynamics that significantly impact strategy adoption and effectiveness.

The framework expands water resources management by integrating AI-driven technologies to simulate realistic decision-making scenarios that reflect both technical requirements and community preferences. This comprehensive approach balances flood damage reduction, water quality improvement, habitat protection, and community recreational needs while incorporating diverse stakeholder viewpoints (Elshaikh-Hayaty et al. 2024; Project Performance International 2024).

Through AI-driven agents representing various stakeholders, the framework facilitates structured dialogues to evaluate mitigation strategies, capturing the complex interplay of technical, social, and economic factors (Salimi et al. 2024; Teague et al. 2022). By simulating decision-making processes across diverse community configurations, the study examines how demographic and socio-economic factors shape public preferences and influence the effectiveness of hydrological actions (Saravi et al. 2019; Yousefi et al. 2020). This approach aligns with social hydrology principles, emphasizing public participation in water resource management (Chang et al. 2023) while fostering community resilience and sustainable environmental planning (Ray 2023).

The MHT framework employs a sophisticated system architecture integrating multiple AI technologies to simulate decision-making processes in water resource management and hazard mitigation. The architecture combines advanced natural language processing (NLP) models with a multi-agent system to create a dynamic, interactive environment for stakeholder collaboration and strategy development.

The framework leverages a multi-agent system where autonomous agents represent diverse community stakeholders, each configured with unique attributes such as age, occupation, and personality. AutoGen manages these structured interactions, ensuring orderly and meaningful dialogues between agents. DarkIdol-Llama-3.1-8B enhances these conversations by supporting extended dialogues and maintaining context over long discussions, enabling complex multi-round negotiations. This flexibility allows for various community configurations, providing insights into how different mitigation strategies may unfold across different contexts.

The study utilizes several AI technologies and large language models (LLMs) including NLP, Autogen, Open AI GPT-4o, and DarkIdol-Llama-3.1-8B for development and implementation of AI agents to the decision-support framework.

Autogen and OpenAI's GPT-4o-mini: AutoGen serves as the backbone for orchestrating multi-agent interactions within the MHT framework, providing the structure necessary for agents to engage in realistic, collaborative dialogues. Each agent, configured with distinct attributes such as age, occupation, and personality, represents a community stakeholder, thus reflecting the diversity typically found in real-world communities (Martelo et al. 2024). AutoGen's flexibility enables these agents to participate in both hierarchical and parallel conversations, simulating the complexity of real-world decision-making, where multiple issues are often deliberated concurrently (Teague et al. 2022). The platform also supports ‘conversation programming,’ which integrates NLP with computational logic, allowing agents to engage in structured negotiations and reconcile competing priorities (Elshaikh-Hayaty Mabrouki & Mohamed 2024).

Enhancing this capability, OpenAI's GPT-4o-mini provides the NLP functionality necessary for generating coherent, contextually relevant responses during multi-agent interactions. While GPT-4o-mini is a more compact version of OpenAI's full-scale model, it retains the essential features needed for simulating detailed and realistic conversations. By interpreting complex language inputs and balancing multiple decision factors, GPT-4o-mini enables agents to adapt their dialogue according to their pre-defined characteristics, such as age, cultural background, and problem-solving styles, thereby increasing the realism of the simulation (Du et al. 2021). The integration between AutoGen and GPT-4o-mini ensures real-time interaction, allowing agents to dynamically adjust their strategies and responses as the conversation progresses, which mirrors the fluid nature of decision-making in real communities (Salimi et al. 2024). These tools provide a comprehensive framework that simulates both the technical and social dimensions of flood mitigation strategies, offering a realistic and data-driven platform for community-level decision-making (Chang et al. 2023).

DarkIdol-Llama-3.1-8B: DarkIdol-Llama-3.1-8B plays a critical role in handling long-context interactions within the MHT framework, ensuring that conversations maintain coherence over multiple rounds of discussion. Developed by QuantFactory, this large-scale language model is optimized for extended dialogues, with the ability to manage up to 35,000 tokens in its context window (QuantFactory 2024). This extended capacity allows agents to refer to earlier decisions and inputs, ensuring that conversations remain consistent and contextually informed as they evolve. DarkIdol's uncensored instruction-based fine-tuning further enhances its capability to engage in detailed, unrestricted dialogues, allowing for comprehensive and realistic discussions during flood mitigation planning (Elshaikh-Hayaty et al. 2024).

Another key feature of DarkIdol is its computational efficiency, particularly its ability to leverage GPU power through torch.cuda, making it ideal for the large-scale simulations used in the MHT framework. This ensures that even with the model's extensive context window and detailed language processing capabilities, it can run smoothly and effectively during resource-intensive simulations (Salimi et al. 2024). By integrating DarkIdol-Llama-3.1-8B with AutoGen and GPT-4o-mini, the MHT framework benefits from its ability to sustain complex, multi-round discussions. This integration allows the simulation to address both short-term and long-term planning, providing deeper insights into how various flood mitigation strategies might unfold over time. DarkIdol's capacity to manage extended, meaningful interactions adds depth to the decision-making process, ensuring that the social and technical implications of water resource management decisions are thoroughly explored and well-understood (Chang et al. 2023).

The framework configures 1,000 agents with diverse attributes reflecting real-world communities (Gent 2024). These attributes shape each agent's decision-making approach and communication style, enabling the simulation to capture a wide array of perspectives. The configuration encompasses demographic information, personality traits, communication preferences, and other characteristics as detailed in Table 1.

Our framework integrates two complementary LLMs to enhance the realism of simulated community-level decision-making. DarkIdol-Llama-3.1-8B provides agents with an extended context window (up to 35,000 tokens), enabling them to maintain coherence across multi-round discussions by referencing earlier decisions and adapting to evolving priorities. GPT-4o-mini complements this by processing immediate contextually rich interactions.

This LLM integration enhances community realism in three key ways: (1) demographic-consistent reasoning, allowing agents to make decisions aligned with their defined attributes rather than following simplistic rules; (2) dynamic knowledge integration, enabling agents to synthesize both technical data and social factors in their deliberations; and (3) natural language negotiation that captures the nuanced dynamics of community deliberation through persuasion and compromise.

The system workflow establishes a structured approach to simulating decision-making processes in water resources planning and hazard mitigation, encompassing initialization, execution, and outcome documentation stages. The simulation begins with an initial message providing agents with comprehensive context about the discussion's objectives. This includes the focus on flood mitigation strategies, primary goals for optimizing resource allocation and protecting water resources, and the available budget of $15 million. The system extracts and compiles essential data from academic papers, project reports, and legal documents using PyMuPDF (fitz) (Artifex Software Inc 2024) python library, creating a structured knowledge base for agent reference during simulations.

Agent Interactions and Decision Process Agents interact under the guidance of a user proxy agent who facilitates discussions and ensures alignment with exercise objectives. The interaction process involves collaborative deliberation on various flood mitigation strategies, considering factors such as cost-effectiveness, long-term sustainability, and public acceptability. Agents engage in structured negotiations, sharing expertise and working toward consensus while representing their assigned stakeholder perspectives. The system meticulously documents all decisions, including selected mitigation options, budget allocations, and the reasoning behind choices. This comprehensive documentation captures both majority decisions and dissenting opinions, providing valuable data for post-simulation analysis.

The framework simulates community decision-making with multiple teams composed of diverse stakeholder agents (Sermet et al. 2020). Each team must allocate a fixed budget of $15 million across local or watershed-wide mitigation options to maximize benefits in flood damage reduction, water quality improvement, habitat protection, and community recreational enhancements. Decisions must be transparently documented, including justifications for strategy selection and fund allocation.

The decision-support framework provides several mitigation alternatives under various categories including capital improvement projects, stormwater management policies, flood protection ordinances and recreational investments.

Capital improvement projects: These projects involve substantial infrastructure investments aimed at controlling and managing water flow to reduce flood risks. Examples include the construction of flood control reservoirs, levees, stormwater detention basins, and channel modifications. These projects typically have high upfront costs but offer significant long-term benefits in terms of flood damage reduction and water management.

Stormwater management policies: These policies focus on the implementation of traditional as well as green infrastructure and other sustainable practices to manage stormwater runoff. Options include the installation of permeable pavements, rain gardens, bioswales, and other low-impact development techniques. These measures are generally more cost-effective than large-scale infrastructure projects and provide benefits such as improved water quality, reduced urban heat islands, and enhanced groundwater recharge.

Flood protection ordinances: These ordinances involve regulatory measures that require new developments and existing structures to incorporate flood protection features. Examples include elevated buildings, flood barriers, and zoning regulations that restrict construction in flood-prone areas. These ordinances help reduce the vulnerability of properties to flood damage and promote long-term resilience in the community.

Recreational investments: These investments aim to enhance community well-being by developing recreational spaces in flood-prone areas, including recreational sport facilities. Options include the creation of parks, trails, sports facilities, playgrounds, and picnic areas. While these investments do not directly reduce flood risks, they provide significant social and economic benefits by improving the quality of life for residents, promoting outdoor activities, and increasing public support for flood mitigation efforts. The associated costs and benefits are carefully considered to maximize the overall value to the community.

By incorporating these diverse mitigation options and adhering to the structured framework rules, the framework provides a robust platform for simulating the decision-making processes using an integrated water resources management approach (Biswas 2004). This comprehensive approach allows stakeholders to explore the complexities and trade-offs of different strategies, fostering a deeper understanding of the local factors that influence effective and socially acceptable water management solutions. Through this interactive and data-driven simulation, the MHT framework aims to contribute to more resilient and sustainable communities by promoting informed and collaborative decision-making.

During initialization, agents are assigned 18 demographic variables and corresponding baseline preferences derived from historical MHT data using correlation matrices. Technical documents are processed to extract engineering constraints, cost–benefit data, and legal considerations. These are organized into structured decision contexts, including goal definitions, rules, and action options.

In the execution phase, AutoGen's GroupChat and GroupChatManager coordinate multi-agent interactions, while DarkIdol-Llama-3.1-8B maintains contextual coherence across two discussion rounds. This phase includes four sub-processes: issue framing, strategy exploration, consensus-building, and budget allocation. Agent conversations are recorded for later analysis.

Data collection occurs continuously, capturing decisions, justifications, and budget distributions. Outputs are stored as CSV files and analyzed through descriptive statistics, predictive modeling, and comparison with human decision patterns. The long-context window of DarkIdol-Llama-3.1-8B enables nuanced tracking of multi-round dialogues, supporting deeper insights into community-driven hazard mitigation planning.

Effective hazard mitigation depends on anticipating future conditions while managing uncertainty in both environmental factors and human behavior. The AI-driven MHT framework integrates future state prediction with uncertainty modeling to support more adaptive and realistic planning. It enables AI agents representing diverse stakeholders to assess the long-term resilience of various mitigation strategies under different climate scenarios. Multi-round decision-making captures how early choices influence later adaptations, producing a dynamic model of evolving community response.

DarkIdol-Llama-3.1-8B's extended context window (35,000 tokens) allows agents to retain memory of past decisions and projected outcomes across simulation rounds. This supports deliberations that weigh both immediate benefits and long-term trade-offs under changing environmental and social conditions. AutoGen further enhances these simulations by coordinating multi-agent interactions, allowing stakeholders to share forecasts, challenge assumptions, and refine strategies collaboratively.

The framework also incorporates a range of uncertainties that affect decision-making. Parametric uncertainties, such as flood frequency and economic loss, are handled through the simulation's multi-scenario approach. Non-parametric uncertainties, like shifting stakeholder preferences, are addressed using context-aware AI processing. Demographic variability among agents introduces internal uncertainty, reflecting realistic differences in community priorities – our results show that younger participants (19–30 years) often prefer immediate infrastructure investments, while older participants (>61 years) favor conservative strategies. External uncertainties, such as climate change and policy shifts, are modeled through the diverse perspectives of the 1,000 simulated agents.

This integrated approach builds on recent advances in AI for predictive planning, such as Tutsoy & Tanrikulu's (2022) demographic-based vaccination strategy model. Unlike prior work, however, the MHT framework incorporates collaborative simulation among diverse stakeholders, offering a broader and more socially grounded platform for hazard mitigation. Combining predictive modeling with uncertainty characterization facilitate delivering a more robust foundation for developing socially acceptable and technically sound mitigation strategies.

The results section presents findings from the MHT framework simulations, examining decision-making processes of diverse community stakeholders through case studies and simulation outcomes. The analysis focuses on strategy selection, budget allocation patterns, and the influence of demographic factors on decision-making.

The MHT framework simulations evaluated how teams allocated a $15,000,000 budget across various mitigation options. These options included capital improvement projects, infiltration policies, freeboard policies, buyouts, and recreational investments, with the goal of maximizing benefits in flood damage reduction, water quality improvement, habitat protection, and community recreational enhancement.

Analysis revealed distinct patterns in strategy selection based on demographic and professional characteristics:

Age-based preferences: The 19–30 age group prioritized immediate infrastructure improvements and capital improvement projects, focusing on transformative actions with immediate benefits. In contrast, participants aged 61–75 years favored long-term, conservative strategies such as buyouts and freeboard policies, emphasizing sustainability and cautious resource management.

Professional background influence: Technical professionals, particularly engineers, showed strong preference for capital-intensive projects, prioritizing robust structural benefits. Those in policy or community-oriented roles favored investments directly enhancing social well-being, such as recreational area development.

Communication and personal factors: Participants with collaborative communication styles and community-focused interests typically advocated for balanced strategies combining infrastructure improvements with recreational investments, demonstrating a holistic approach to community resilience.

The simulation results highlighted several critical factors influencing water resource management decisions:

Demographic impacts: Age emerged as a primary determinant in budget allocation, with younger participants favoring immediate infrastructure investments while older participants preferred conservative spending strategies focused on long-term sustainability.

Professional expertise: Occupation significantly influenced strategy selection. Technical experts consistently prioritized capital-intensive projects, while those from policy-making or community service backgrounds emphasized social well-being enhancements.

Social dynamics: Communication styles and personal interests shaped by decision-making beyond technical considerations. The simulations revealed that interpersonal dynamics and personal values played crucial roles in strategy development and implementation.

Additional factors such as relationship status, language proficiency, and cultural background, while secondary to age and occupation, contributed to the complexity of budget allocation decisions. These findings emphasize the importance of flexible, inclusive decision-making frameworks that can integrate diverse perspectives for optimal outcomes.

The analysis of the simulation outcomes provides a comprehensive understanding of the decision-making processes and the effectiveness of various strategies employed by different teams. The metrics for evaluating the effectiveness of the strategies include budget allocation, option selection, and the impact of various features on these decisions. The following charts illustrate these metrics and their interpretations.

The MHT framework, supported by AI, reveals how different demographic factors influence budget allocation in flood mitigation. Age emerged as the most significant factor, with older participants (61–75 and >76 years) adopting a conservative approach, reserving more funds for future contingencies. This indicates a preference for long-term sustainability. In contrast, younger participants (19–30 years) spent heavily on immediate infrastructure improvements, reflecting a proactive mindset and a higher risk tolerance, focusing on current needs over future uncertainties.

Occupation also plays a key role in shaping budget decisions. Participants with technical backgrounds, such as engineers, allocated more funds to capital-intensive projects, emphasizing long-term structural benefits. Meanwhile, participants in policy-making or community-oriented roles prioritized regulatory measures and social investments, such as recreational areas and public health initiatives. These differences reflect the contrasting priorities of technical versus community-focused participants.

Communication style and personal interests further influence spending behavior. Collaborative participants with a strong interest in community-centered investments advocated for balanced spending on both recreational add-ons and infrastructure, indicating that personal values and social engagement shaped their decision-making processes. Socio-demographic factors, including language proficiency, relationship status, and cultural background, added complexity to budget decisions. For example, participants with higher language proficiency negotiated budgets more effectively, while those with diverse cultural backgrounds prioritized different strategies based on their perceptions of risk and resilience.

Figure 2 illustrates the relative importance of demographic factors in budget allocation decisions, with age emerging as the most influential parameter (importance value: 0.16). The visualization quantifies how each attribute affects spending patterns in flood mitigation planning, creating a clear hierarchy of influence. Occupation represents the second most significant factor (importance value: 0.10), followed by interests and hobbies, communication style, and values/beliefs. The substantial gap between age and other factors visually emphasizes its outsized role in determining resource allocation preferences. This data-driven prioritization provides policymakers with a road map for anticipating community spending priorities based on demographic profiles, enabling more targeted stakeholder engagement strategies. Notably, factors like gender and preferred communication medium show comparatively minimal influence, suggesting these characteristics play secondary roles in budget decisions.

Occupation further impacted strategy selection. Engineers and technical professionals opted for infrastructure-heavy solutions like channel modifications and capital projects, while those in policy-making roles emphasized social and community-based options, such as recreational developments and buyouts. This divide underscores the influence of professional background on what participants perceive as effective mitigation. Values and beliefs also shaped decisions, with environmentally focused participants supporting strategies that enhance environmental quality or social cohesion, such as infiltration policies and recreational investments. This demonstrates how personal ethics and priorities influence mitigation choices.

Communication style and personal interests influenced holistic decision-making. Participants who favored collaboration selected a balanced mix of infrastructure and social investments, reflecting their preference for teamwork and consensus-building. Similarly, participants with outdoor or environmental interests were more likely to support recreational area development. Other socio-demographic factors like cultural background, gender, and language proficiency also played a role, albeit to a lesser extent. For example, participants with specific cultural views on risk tended to prefer strategies aligned with their cultural understanding of resilience. Figure 3 illustrates these diverse preferences.

Figure 3 presents a quantitative analysis of the demographic factors influencing mitigation option selection, with age again emerging as the dominant parameter (importance value: 0.14). The visualization reveals how different personal attributes shape strategy preferences, creating a comprehensive decision model. Occupation appears as the second most influential factor (importance value: 0.08), followed by interests and hobbies and cultural background. This hierarchical arrangement provides planners with clear guidance on which demographic factors most strongly predict strategy selection tendencies. The results demonstrate that while budget allocation (Figure 2) and option selection (Figure 3) share common influential factors, the relative importance of secondary factors shifts between these two decision dimensions. This nuanced understanding enables more targeted community engagement strategies that address the specific demographic drivers of different aspects of the decision-making process.

Interestingly, a moderate correlation (0.44) between educational background and cultural background suggests that individuals with similar educational levels tend to share cultural similarities, which may influence their decision-making preferences. On the other hand, weak correlations, such as the one between language proficiency and technology savviness (–0.10), indicate that higher language skills don't necessarily align with higher technical skills, highlighting the diversity in participant backgrounds. This heatmap provides valuable insights into how these variables interact, informing future planning and flood mitigation efforts by offering a more nuanced understanding of the factors driving decision-making.

Across both models, participants aged 19–30 years showed a strong preference for Buyouts and Recreational Area Development, reflecting their focus on immediate relief and community enhancement. In Figure 6 (GPT-4o-mini model), this pattern remains consistent, with Buyouts showing the highest frequency (approximately 175 selections) for the youngest age group, followed closely by Recreational Investments.

The 31–45 age group exhibited similar interests to the younger cohort, but with an increased emphasis on Freeboard Policy, suggesting a desire to balance immediate needs with future precautions. As shown in Figure 5, this age group's Freeboard Policy selections increased to approximately 150 instances, nearly matching their Buyouts frequency, demonstrating a more balanced approach between immediate and long-term strategies.

Those in the 46–60 age group also emphasized Buyouts and Recreational Area Development, indicating widespread support for these measures. Both figures show that Infiltration Policy selections remained relatively consistent across all age groups at approximately 150–175 instances, suggesting this option serves as a universal baseline strategy regardless of age.

Participants aged 61 years and above tended to prioritize long-term strategies such as the Freeboard Policy, although Buyouts remained a popular choice across all age groups. In Figure 5, the >76 age group shows nearly equal preference for Buyouts, Freeboard Policy, and Infiltration Policy (all around 125–150 selections), indicating a more diversified approach to risk management. Notably, Capital Improvement Projects remain the least selected option across all age groups in both models, with frequencies rarely exceeding 50 selections.

This distribution of preferences suggests that while younger participants are inclined toward immediate actions, older groups favor more cautious, long-term approaches. The comparative analysis between Figures 5 and 6 reveals that while both models capture similar age-based trends, DarkIdol-Llama-3.1-8B shows slightly higher frequencies for recreational investments across all age groups, aligning with its overall emphasis on community-centered enhancements.

Similarly, the female group exhibits a preference for buyouts and recreational investments, with a focus on reducing current flood risks and enhancing community spaces. In contrast, the male group demonstrates mirrored preferences, also choosing buyouts and current traditional CIP costs, but placing less emphasis on future-oriented strategies like freeboard policy. These gender-specific patterns indicate consistent priorities across all groups, with a shared emphasis on immediate actions and community enhancements.

In contrast, the DarkIdol-Llama-3.1-8B model displays a greater preference for larger, infrastructure-heavy facilities like trails, playgrounds, and shelters, indicating a focus on long-term investment. This approach targets the development of durable, multi-functional spaces that serve both recreational and environmental objectives. The differences between the models underscore a balance between short-term community engagement and long-term infrastructure investment, providing complementary insights into recreational priorities across the planning strategies.

The discussion section interprets the findings from the MHT framework simulations, focusing on the strengths and limitations of the approach and its broader implications for water resource management and hazard mitigation. By leveraging AI-driven tools, the framework provides a comprehensive analysis of the factors influencing decision-making processes. This section explores key insights from the study, highlights the strengths and limitations of the MHT framework, and provides recommendations for improving the approach and its real-world applications.

The MHT framework, enhanced by AI-driven tools, has provided useful insights into decision-making processes for water resource management and hazard mitigation, capturing diverse scenarios and a range of demographic perspectives. However, contrary to expectations, there were minimal differences in the options chosen across age groups. Most participants selected similar strategies with only small variances, primarily in preferences for recreational investments. Specifically, DarkIdol-Llama-3.1-8B demonstrated a unique emphasis on recreational investments, aligning with community-centric goals and quality-of-life improvements in flood resilience.

This limited variance across demographics may reflect the challenges current LLMs face in making sophisticated, multi-step connections between persona attributes and situation-specific behaviors. While the models simulate basic preferences, they struggle to fully capture complex, persona-driven behaviors in extended, multi-stage decision processes. Additionally, with a substantial budget and few available options, agents tend toward larger selections to maximize spending, rather than conserving resources. Thus, the real differentiation in decision-making emerges in the types of options chosen rather than the overall amount spent, underscoring the need for more nuanced decision models.

The MHT framework's primary strength is its ability to simulate a broad range of decision-making scenarios, accounting for diverse individual characteristics. The integration of AI tools enables advanced language processing and strategic modeling, offering insight into how age, occupation, and values influence flood mitigation choices. For instance, the recreation investment analysis revealed slight distinctions, with younger participants and certain groups, like the genderqueer and non-disclosed, favoring small-scale recreational investments. DarkIdol-Llama-3.1-8B, on the other hand, showed preferences for larger, community-enhancing recreational projects, highlighting varied perspectives on long-term investments. Another strength is the MHT framework's collaborative environment, where agents representing different community stakeholders engage in structured discussions. This interaction fosters more inclusive and balanced flood mitigation strategies, ensuring that diverse viewpoints are incorporated.

Despite its strengths, the framework has limitations. A key issue is the reliance on static agent characteristics, which restricts flexibility. While models like DarkIdol-Llama-3.1-8B and GPT-4o-mini aid in nuanced decision-making, they cannot fully capture dynamic human decision-making, which evolves with emotional responses, social pressures, and real-time updates. The framework's rigid structure thus limits its ability to mirror complex, adaptive decision environments accurately. Additionally, the framework's assumptions and pre-defined agent roles may introduce bias, particularly in scenarios requiring spontaneous adaptation to evolving contexts. This lack of flexibility in agent roles constrains the model's realism, as human decisions are often context-driven and highly adaptive.

Performance evaluation of the framework reveals additional limitations. The simulation results showed a minimal variance in option selection across different demographic groups, with most participants choosing similar strategies regardless of their assigned attributes. This homogeneity suggests that current language models struggle to create meaningful persona-based behavioral differentiation in complex multi-step decision processes. Furthermore, with a substantial budget of $15 million and limited mitigation options, agents consistently selected most available options to maximize budget utilization, resulting in decision differentiation occurring only in the specific types of options chosen rather than in strategic diversity. This pattern indicates that the framework's performance is more sensitive to budget constraints and available options than to the nuanced demographic characteristics of agents.

Another challenge is the requirement for extensive data and computational power. The high computational demands for running models like DarkIdol-Llama-3.1-8B can make the framework less accessible to small communities or organizations with limited resources. The implementation of these models involves substantial data processing capabilities, which may pose a barrier to adoption in regions with constrained technological infrastructure. Developing more lightweight or optimized versions of these models could help mitigate this issue and improve accessibility.

The simulations emphasize the importance of inclusive decision-making that accounts for diverse community preferences. Using both DarkIdol-Llama-3.1-8B and GPT-4o-mini, the results demonstrate the need to balance immediate flood protection with long-term structural improvements. Future iterations of the MHT framework would benefit from dynamic agent configurations that adapt to new information and evolving circumstances, better reflecting real-world decision-making. Incorporating broader behavioral models and real-world community input could reduce potential biases while simplifying the interface and providing educational resources would make the framework more accessible to smaller communities and non-experts.

To overcome extensive data and computational power limitations, several optimal solutions can be implemented. First, model compression techniques such as quantization and knowledge distillation could reduce the computational requirements of LLMs while maintaining performance quality. For instance, deploying 4-bit or 8-bit quantized versions of DarkIdol-Llama-3.1-8B could reduce memory requirements by up to 75% with minimal accuracy loss. Second, implementing a cloud-based service architecture with API access would allow resource-constrained organizations to utilize the framework without local computational infrastructure. Third, developing a hierarchical processing approach where initial screening uses lightweight models (like GPT-4o-mini) and only complex scenarios invoke resource-intensive models could optimize computational efficiency. Fourth, creating pre-computed decision templates based on common community profiles could reduce real-time processing needs. Finally, establishing regional computational hubs or partnering with academic institutions could provide shared resources for smaller communities, making advanced AI-driven decision-support accessible regardless of local technological constraints.

The insights from the MHT framework underscore valuable applications for real-world water resources planning and hazard mitigation. Through a thorough analysis of how demographic and socio-economic factors influence decision-making, the framework enables strategies that are both technically robust and representative of community needs. Policymakers can leverage the MHT's data-driven insights for developing flood mitigation strategies that balance short-term actions with long-term goals, directly support policy development, resource allocation, and stakeholder engagement. This contributes to outcomes such as improved flood damage reduction, water quality, community resilience, and recreational benefits.

The MHT framework's ability to simulate diverse decision-making scenarios, incorporating DarkIdol-Llama-3.1-8B for complex, long-term strategies and GPT-4o-mini for immediate actions, reveals significant real-world applications. For instance, age and gender groups were observed to prioritize a range of options, from buyouts to longer-term infrastructure investments like freeboard policies. These findings illustrate the value of inclusive flood mitigation strategies that reflect a broad spectrum of community preferences.

Moreover, the MHT framework's insights can help policymakers design flood protection strategies that integrate immediate actions, such as buyouts, with resilient infrastructure investments. The data-driven approach also emphasizes the need to engage diverse stakeholders, as their perspectives contribute to more balanced and widely supported flood mitigation efforts. This approach enables policymakers to pinpoint the most impactful measures, making resource allocation more efficient and maximizing community benefits.

To fully realize the potential of the MHT framework, policymakers and practitioners should prioritize inclusive, participatory decision-making processes. The simulation results demonstrate that incorporating diverse community perspectives leads to more representative and sustainable strategies. Using AI tools allows decision-makers to capture the nuanced preferences of stakeholders, aligning strategies with both short-term and long-term community objectives.

Policymakers should also emphasize collaborative planning, involving technical experts, community leaders, and residents. This approach ensures that all relevant viewpoints contribute to the development of comprehensive mitigation strategies. Additionally, improving the framework's accessibility and usability through user-friendly interfaces and training resources can extend its benefits to smaller organizations and communities with limited resources.

Future development should focus on enhancing the adaptability and usability of the MHT framework to better mirror real-world decision-making. Although current models offer substantial potential, additional improvements are needed to incorporate adaptive behavioral models that evolve with real-time data and shifting circumstances. A promising area for further research includes integrating dynamic agent characteristics that adjust in response to new information and stakeholder input, allowing simulations to better reflect the fluid nature of decision-making.

Expanding the MHT framework to analyze scenarios related to climate change, urbanization, and population growth would provide deeper insights into how emerging challenges may influence decision-making. Exploring behavioral models that account for psychological and emotional influences could enhance the realism of the simulations, offering a fuller picture of how communities respond to flood mitigation strategies over time. Longitudinal studies could also track decision impacts, adding depth to the MHT's predictive capability.

Additionally, improving the framework's accessibility and scalability will be essential. Interactive dashboards, visualization tools, and user-friendly interfaces can help non-experts more easily interpret results. Incorporating real-time environmental data, community feedback, and social media analytics can increase simulation accuracy and relevance, making the framework even more effective for tailored decision-making.

Extending the MHT framework as an interactive, web-based simulation tool could significantly benefit the hydrology and water management fields. Such an environment would allow real-time simulation of different goals and decision-making tasks, enabling users to instantly observe how various community segments might behave or support specific options under given scenarios. This interactive component would be invaluable for both public education and expert analysis, providing a visual and data-rich platform for exploring the social dynamics of decision-making.

This study utilized the MHT framework to simulate decision-making processes in water resource management and hazard mitigation, exploring how factors like age, gender, and occupation influence flood mitigation strategies and resource allocation. By simulating diverse scenarios with AI agents representing various community stakeholders, the study provided insights into demographic influences on decision-making. Two advanced AI models, DarkIdol-Llama-3.1-8B and GPT-4o-mini, supported a nuanced analysis, revealing the priorities of different demographic groups.

Notably, the study found minimal differences in selected options across age groups, with most participants choosing similar strategies. However, DarkIdol-Llama-3.1-8B notably emphasized recreational investments, highlighting a focus on community-centered enhancements alongside flood resilience. The results showed that present LLMs are limited in making sophisticated, multi-step connections between an agent's persona and its behavior in specific situations. The large budget in the simulation led to high option selection rates, as most roles aimed to utilize their allotted funds, resulting in only slight differences in decision-making across demographics. The true variety was observed in the specific options chosen, influenced by the variety of available choices.

This study illustrates the transformative potential of AI in enhancing decision-making frameworks for water resource planning and hazard mitigation. By fostering collaborative decision-making among AI agents configured as diverse community stakeholders, the MHT framework enabled a more comprehensive understanding of community needs and promoted strategies that gain broad-based support, ultimately enhancing resilience.

The integration of AI-driven agents into water resource management has promising implications for real-world applications. Modeling diverse stakeholder perspectives helps policymakers develop strategies that are inclusive, adaptable, and effective. AI tools also enable the rapid processing of extensive datasets, making it possible to evaluate complex, evolving scenarios in real-time. This adaptability is particularly valuable for addressing dynamic challenges like climate change, urbanization, and population growth, where flexible, responsive planning is critical.

Practically, AI-driven simulations can guide urban planning, infrastructure development, and disaster risk reduction by providing data-driven insights tailored to specific community requirements. Policymakers and practitioners can use these tools to assess the immediate and long-term effects of interventions, ensuring that strategies are not only technically sound but also socially responsive. In conclusion, integrating AI into decision-making frameworks like the MHT offers critical insights for creating sustainable, resilient, and inclusive water resource management strategies. As AI technology advances, its role in policy development and resource management will become increasingly crucial in equipping communities to face future environmental challenges and uncertainties.

During the preparation of this manuscript, the authors used ChatGPT, based on the GPT-4 model, to improve the flow of the text, correct grammatical errors, and enhance the clarity of the writing. The language model was not used to generate content, citations, or verify facts. After using this tool, the authors thoroughly reviewed and edited the content to ensure accuracy, validity, and originality, and take full responsibility for the final version of the manuscript.

This project was funded by the National Oceanic and Atmospheric Administration (NOAA) via a cooperative agreement with the University of Alabama (NA22NWS4320003) awarded to the Cooperative Institute for Research to Operations in Hydrology (CIROH).

L.K. contributed to conceptualization, methodology, software, validation, formal analysis, investigation, data curation, writing – original draft, and visualization. R.S. contributed to validation, writing – review & editing, and data curation. Y.S. contributed to conceptualization, methodology, writing – review & editing, investigation, validation, funding acquisition, and visualization. M.M. contributed to resources, writing – review & editing. I.D. contributed to conceptualization, methodology, writing – review & editing, project administration, supervision, funding acquisition, and resources.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.