As a new year begins, it is appropriate to reflect on the evolving role of modelling in the water sector. There is growing recognition that traditional approaches—optimised for bounded systems, design conditions, and steady-state assumptions—are increasingly challenged by changing environmental drivers, operational volatility, and the expansion of available data sources. In response, four developments are reshaping expectations of how models are conceived, deployed, and maintained:
- integration of environmental and external data to improve situational relevance,
- scenario-based modelling for crisis and resilience planning under non-stationary conditions,
- localisation and context-aware configuration to sustain validity across diverse operating regimes, and
- the deployment of next-generation AI and machine learning as mechanisms for adaptation, data fusion, and lifecycle model governance.
These developments do not imply a rejection of established methods. Instead, they reflect a gradual shift from asset-bounded, static modelling practices toward dynamic modelling frameworks that engage directly with environmental conditions, simulate extreme events to support preparedness, maintain alignment with local operational context, and incorporate AI-enabled feedback mechanisms to mitigate performance drift over time. While adoption remains uneven, the trajectory suggests that relevance, resilience, and adaptability will become core performance criteria for modelling systems moving forward.
1. Integration of open environmental data into water models
The growing availability of environmental, infrastructural, and socio-technical data has shifted expectations for modelling in the water sector. Models are no longer viewed as isolated analytical tools, but as components of broader knowledge systems that draw on diverse data sources to represent real-world conditions more accurately. This shift is central to current discussions on the integration of data and models.
Early efforts to pursue integration can be traced to the development of frameworks linking sewers, treatment processes, hydrology, and receiving waters (e.g., Vanrolleghem et al., 1998; Rauch et al., 2002). However, widespread deployment remained constrained by interoperability limitations, fragmented data availability, and the computational demands of running multi-domain simulations at relevant temporal and spatial scales. Surrogate modelling (e.g., Meirlaen et al., 2002) emerged as a strategy to reduce runtime by approximating complex system behaviour, but operational uptake was limited due to non-linear response behaviour, reduced validity outside calibration domains, and the high sampling requirements needed to maintain fidelity.
Recent progress has been driven less by conceptual novelty and more by enabling conditions: improved computational capacity, evolving data standards, and architectures designed explicitly for interoperability. A notable example is the Meta-Scientific-Modeling (MSM) framework, proposed as a sustainable and accessible approach to both open data and open modelling integration (e.g.: Sala et al., 2020). MSM introduces an Open Data architecture that systematically manages heterogeneous external datasets—including environmental, infrastructural, and socio-economic inputs—through modular data agents responsible for accessing, structuring, and harmonising data from remote and inconsistent sources. Complementing this, an Open Modeling architecture enables the coupling of heterogeneous mechanistic models without bespoke interface development by standardising interaction protocols and using model agents to manage communication. Together, these features illustrate how integration can be operationalised rather than merely theorised.
Practical demonstrations of this shift were visible at the recent IWA- Urban Drainage Modelling Conference UDM2025 Workshop 4 (Influent Flowrate Generators for System Design in Urban Drainage Modelling using Open Data Sources), which showed how internet-accessible datasets—meteorological records, digital elevation models, topographic layers, land-use information, and urban growth projections—can be embedded into hydraulic models to estimate sewer flowrates and support early-stage design. These workflows represent a step toward catchment-aware modelling, where external drivers are treated as integral model inputs rather than contextual annotations.
As data access improves and integration architectures mature, the primary challenges are shifting from technical feasibility to data governance, quality assurance, and validation under uncertainty, especially where models inform planning or investment decisions. Integration is best understood not as a replacement of existing models, but as a means of making them more contextually representative, operationally relevant, and decision-ready.
2. Modelling for crisis simulation and resilience planning
Traditional modelling practices have largely centred on assessing system performance under nominal or design conditions. The growing influence of climate variability, energy supply instability, and emerging systemic risks has shifted attention toward understanding performance under degraded or disrupted states. As a result, there is increasing interest in prospective, scenario-based approaches that evaluate system response to credible disturbances rather than relying solely on retrospective failure analysis.
Applications of Model-Based Systems Engineering (MBSE) and system-of-systems modelling illustrate this shift. These approaches map failure propagation pathways, assess the effectiveness of intervention strategies, quantify bottlenecks and redundancy limits, and estimate recovery timelines (Joannou et al., 2019). However, their practical deployment requires context-specific data, skilled interpretation, and governance frameworks before routine operationalisation is feasible (Juan-Garcia et al., 2021).
In parallel, digital twins configured for stress-testing — crisis twins —aim to support preparedness and decision rehearsal by simulating degraded operating conditions, energy constraints, industrial discharge anomalies, and contamination sequences. These systems remain experimental, but indicate a move toward using models not only for optimisation but for resilience planning and uncertainty navigation.
Crisis modelling is thus best understood as a complement to conventional modelling rather than a replacement: a second lens that addresses performance where historical calibration offers limited guidance.
3. Localised and context-aware modelling
Accuracy in water system modelling is increasingly constrained not by methodological limits, but by contextual divergence between the model and the system it represents. Both knowledge-based (mechanistic) models and data-driven models degrade when real-world conditions deviate from their design assumptions.
For data-driven systems, this manifests as concept drift, where statistical relationships no longer hold. For mechanistic systems, the equivalent is structural drift, where boundary conditions and simplifying assumptions no longer match current operations. The result is the same: the model becomes progressively disconnected from the system, even if technically “correct” within its original design space.
Effective contextualisation of local and regional conditions therefore requires more than parameter adjustment; it requires continuous alignment mechanisms. In practice, this implies feedback loops between digital and physical systems—adaptive calibration routines, periodic retraining, boundary condition revision, and validation checkpoints. Without these mechanisms, models become static 'artefacts'. With them, models function as living representations capable of sustaining relevance despite non-stationarity and changing conditions.
In this light, staying in tune with local context is not an optional refinement but a prerequisite for maintaining validity in distributed or multi-site deployments.
4. Deployment of next-generation AI and machine learning models
The role of AI in water modelling is shifting from experimental enhancement to operational infrastructure. As data heterogeneity, non-stationarity, and cross-domain dependencies increase, AI and machine learning are becoming essential for enabling models to function under contemporary conditions.
AI contributes where manual workflows or static models struggle:
- interpreting multimodal data streams,
- detecting divergence and performance drift,
- supporting recalibration under changing boundary conditions,
- automating scenario synthesis for resilience studies, and
- translating analytical outputs into decision-relevant guidance.
In this landscape, hybrid modelling (e.g.: Serrao et al., 2024) could have technical and strategical significance. Instead of positioning data-driven systems against mechanistic or deterministic models, hybrid approaches assign each methodology to the domain where it is structurally strongest: Mechanistic models maintain physical (mass balance) coherence, interpretability, and regulatory alignment, while AI components provide adaptive capacity, closure of unknowns, and continuous tuning. This reflects a division of cognitive labour: physics for structure and causality, AI for inference and adaptation.
Generative AI techniques are promising to extend these capabilities. Variational autoencoders, diffusion models, and sequence-generating architectures can support modelling workflows by synthesising credible scenario variants, augmenting sparse datasets, reconstructing missing data intervals, and generating edge-case events for resilience testing. In this sense, generative AI methods will further facilitate a hybrid modelling approach. Their value lies not in replacing hydraulic or process simulation models, but in predicting system behaviour under conditions where empirical data is sparse or unavailable.
Next-generation AI will have the greatest value when embedded into the model lifecycle, not stacked beside it: monitoring, adjustment, governance, and transparent revision become part of modelling practice rather than exceptions.
Concluding remarks
Recent developments highlight both the limitations of established modelling practices and the multiple pathways available for progress. Rather than advocating for the replacement of existing methods, the most credible advances are emerging through selective integration: external data to improve situational relevance, scenario-based modelling to inform resilience planning, connection to local states to sustain validity across operating regimes, and AI-enabled lifecycle management to maintain alignment under changing conditions.
The challenge ahead is less about methodological novelty and more about ensuring that modelling choices remain proportionate, transparent, and operationally grounded. The effectiveness of next-generation modelling will depend on the maturity of mechanisms that maintain alignment between digital representations and physical systems—continuous calibration, adaptive configuration, and feedback loops compatible with both mechanistic and data-driven components.
Organisations that prioritise relevance and resilience alongside accuracy, and that treat models as dynamically evolving assets, will be best positioned to benefit from ongoing innovation. The next phase of modelling progress is therefore unlikely to be defined by new paradigms alone, but by models designed to remain valid as conditions change.
