logseq/pages/@Modeling AI-Driven Workflows for Ecosystem Resilience Prediction.md

tags:: #zotero title:: @Modeling AI-Driven Workflows for Ecosystem Resilience Prediction item-type:: journalArticle original-title:: Modeling AI-Driven Workflows for Ecosystem Resilience Prediction language:: en authors:: Tiago Sousa, Nicolas Guelfi, Benoit Ries library-catalog:: Zotero links:: Local library, Web library

• Abstract
  • Engineering AI-driven workflows to predict ecosystem resilience is increasingly important for responding to rapid environmental change. As climate variability intensifies, leveraging artificial intelligence to support ecological decision-making has become a common approach. However, prevailing approaches to engineering these prediction systems are often ad-hoc and difficult to reproduce, lacking a systematic methodology to ensure coherence between ecological requirements, data, and AI engineered components. To address this gap, this paper introduces a model-driven engineering methodology to support the systematic development of Ecosystem Resilience Prediction Systems (ERPS). We propose an adaptive workflow that formalizes our development process, based on two core metamodels: EcoSys, for structuring the ecological problem domain by defining its measurable properties as resilience indicators, and PredSys, for defining the AI and dataset solution domains. This approach ensures that AI architectures and datasets are co-designed to predict these properties, formally linking them to ecological requirements. The workflow’s flexible design accommodates goal-driven, datadriven, and model-driven entry points. We demonstrate the methodology’s applicability by retrospectively modeling two realworld case studies from the literature: one on forest carbon stock estimation and another on coastal fisheries management.

• Attachments

  • PDF {{zotero-imported-file Y4XBT2Y6, "Sousa et al. - Modeling AI-Driven Workflows for Ecosystem Resilience Prediction.pdf"}}

• Notes

  • Annotations

    (29/7/2025, 15:12:27)

    • “AI-Driven Workflows for Ecosystem Resilience Prediction” (Sousa et al., p. 1) #5fb236

  - “As climate variability intensifies, leveraging artificial intelligence to support ecological decision-making has become a common approach” (Sousa et al., p. 1) #5fb236
  * *
  
   
  
  - “model-driven engineering methodology to support the systematic development of Ecosystem Resilience Prediction Systems (ERPS).” (Sousa et al., p. 1) #ffd400
  *What do you mean with ecosystem resilience? *
  
   
  
  - “This approach ensures that AI architectures and datasets are co-designed to predict these properties, formally linking them to ecological requirements.” (Sousa et al., p. 1) #5fb236
  * *
  
   
  
  - “one on forest carbon stock estimation” (Sousa et al., p. 1) #5fb236
  * *
  
   
  
  - “nother on coastal fisheries management.” (Sousa et al., p. 1) #5fb236
  * *
  
   
  
  - “ecosystem resilience” (Sousa et al., p. 1) #ffd400
  *natural ecosystems? what? you need to define it early in the paper. *
  
   
  
  - “While modern ecological monitoring generates vast and heterogeneous datasets from automated sensors and large-scale data acquisition [3], applying advanced AI to this data is essential to enabling the repeatable ecosystem assessments needed to track environmental shifts and inform policy” (Sousa et al., p. 1) #5fb236
  * *
  
   
  
  - “despite the potential of AI, the primary engineering barrier is the overall workflow rather than individual models.” (Sousa et al., p. 1) #5fb236
  * *
  
   
  
  - “This absence of systematic and reproducible workflows slows scientific progress and the development of scalable solutions” (Sousa et al., p. 1) #5fb236
  * *
  
   
  
  - “Ecosystem Resilience Prediction Systems (ERPS)” (Sousa et al., p. 1) #5fb236
  * *
  
   
  
  - “The methodology is founded on two core, formally defined metamodels that structure the problem and solution domains.” (Sousa et al., p. 1) #5fb236
  * *
  
   
  
  - “Ecosystem metamodel (EcoSys” (Sousa et al., p. 1) #2ea8e5
  * *
  
   
  
  - “Prediction System metamodel (PredSys)” (Sousa et al., p. 1) #2ea8e5
  * *
  
   
  
  - “AI architectures (PredSys-AI) and dataset characteristics (PredSys-Dataset)” (Sousa et al., p. 1) #2ea8e5
  * *
  
   
  
  - “workflow that accommodates goal-driven, data-driven, and model-driven design processes” (Sousa et al., p. 1) #2ea8e5
  * *
  
   
  
  - “modeling of bidirectional links between ecosystem requirements, data constraints, and AI capabilities” (Sousa et al., p. 1) #5fb236
  * *
  
   
  
  - “coherent” (Sousa et al., p. 1) #ffd400
  *what do you mean? Data constraints on what? *
  
   
  
  - “ecosystem science and prediction systems engineering” (Sousa et al., p. 1) #5fb236
  * *
  
   
  
  - “A central problem in modern ecosystem science is the transition from creating isolated models to engineering integrated, end-to-end workflows.” (Sousa et al., p. 2) #ffd400
  *It is necessary to give some explanatory examples early in the paper just to clearly state what's the focus/context of the work. *
  
   
  
  - “A major concern is ensuring that as AI becomes more powerful, systems remain transparent, accountable, and subject to expert oversight.” (Sousa et al., p. 2) #a28ae5
  * *
  
   
  
  - “While this highlevel, strategic perspective calls for a formal, human-driven design process, it does not prescribe a concrete methodology for achieving it, which is a primary motivation for our work.” (Sousa et al., p. 2) #5fb236
  * *
  
   
  
  - “Another key challenge is the automation and standardization of the data-to-model lifecycle.” (Sousa et al., p. 2) #5fb236
  * *
  
   
  
  - “They highlight the need for automation but do not provide a formal design methodology to specify what these automated pipelines should do and why.” (Sousa et al., p. 2) #5fb236
  * *
  
   
  
  - “This gap is further evidenced by the proliferation of specific, applicationoriented workflows, for instance, for monitoring nocturnal insects [12], managing coastal fisheries [13], and assessing algal blooms [14].” (Sousa et al., p. 2) #ffd400
  *Only at this point you understand that the paper is about natural ecosystems. *
  
   
  
  - “prediction system” (Sousa et al., p. 2) #5fb236
  * *
  
   
  
  - “This allows for precise structural descriptions but offers less support for specifying the model’s requirements or expected behavior” (Sousa et al., p. 2) #5fb236
  * *
  
   
  
  - “De la Vega et al. [18] present a DSL to automate the transformation of structured data into a format suitable for prediction systems.” (Sousa et al., p. 2) #5fb236
  * *
  
   
  
  - “a posteriori documentation of existing datasets” (Sousa et al., p. 2) #5fb236
  * *
  
   
  
  - “few provide the level of detail and guidance needed to model the specific activities, roles, and dependencies involved in engineering prediction systems.” (Sousa et al., p. 2) #a28ae5
  * *
  
   
  
  - “tool concept that provides step-by-step support for explicitly modeling ML application integration in BPMN business processes.” (Sousa et al., p. 2) #a28ae5
  * *
  
   
  
  - “a DSL to model AI engineering processes, covering the needs for such processes as described in academic and industry proposals” (Sousa et al., p. 2) #e56eee
  * *
  
   
  
  - “DSL for modeling ML engineering processes, including activities, their orchestration, roles, resources, and artifacts.” (Sousa et al., p. 3) #5fb236
  * *
  
   
  
  - “Although comprehensive, their approach remains primarily descriptive, supporting the documentation and execution of workflows but lacking mechanisms to enforce consistency between the generated artifacts through verifiable means.” (Sousa et al., p. 3) #5fb236
  * *
  
   
  
  - “In contrast to prior work, we emphasize the formal alignment between problem specification and model design, the derivation of AI and dataset requirements from domain-level concerns, and a prescriptive process model that enforces consistency across artifacts.” (Sousa et al., p. 3) #e56eee
  * *
  
   
  
  - “AIdriven workflows for ecosystem resilience prediction.” (Sousa et al., p. 3) #a28ae5
  * *
  
   
  
  - “Ecosystem metamodel (EcoSys)” (Sousa et al., p. 3) #2ea8e5
  * *
  
   
  
  - “Prediction System metamodel (PredSys)” (Sousa et al., p. 3) #2ea8e5
  * *
  
   
  
  - “AI models and datasets.” (Sousa et al., p. 3) #2ea8e5
  * *
  
   
  
  - “adaptive model-driven workflow.” (Sousa et al., p. 3) #5fb236
  * *
  
   
  
  - “The section concludes with three detailed scenarios that showcase the adaptability of our approach in real-world contexts.” (Sousa et al., p. 3) #5fb236
  * *
  
   
  
  - “A. Metamodeling Requirements, Data, and AI Components” (Sousa et al., p. 3) #2ea8e5
  * *
  
   
  
  - “ERPS” (Sousa et al., p. 3) #5fb236
  * *
  
   
  
  - “Ecosystem metamodel (EcoSys)” (Sousa et al., p. 3) #5fb236
  * *
  
   
  
  - “Prediction System metamodel (PredSys)” (Sousa et al., p. 3) #5fb236
  * *
  
   
  
  - “1) Ecosystem metamodel (EcoSys)” (Sousa et al., p. 3) #2ea8e5
  * *
  
   
  
  - “Our ecosystem metamodel (EcoSys), inspired by the DREF framework [5], adapts foundational resilience concepts from software engineering to natural ecosystems (see Figure 1).” (Sousa et al., p. 3) #5fb236
  * *
  
   
  
  - “s :” (Sousa et al., p. 3) #ff6666
  * *
  
   
  
  - “Evolution axes define how these entities or the ecosystem health change over an ordered dimension (e.g. time).” (Sousa et al., p. 3) #5fb236
  * *
  
   
  
  - “resilience patterns” (Sousa et al., p. 3) #ffd400
  *you need to give some examples. What do you want predict? *
  
   
  
  - “2) Prediction System metamodel - AI part” (Sousa et al., p. 4) #2ea8e5
  * *
  
   
  
  - “Furthermore, since neural networks contain multiple components in its structure, either in a linear or non-linear order, we define clEdge and clOrdering classes to allow precise specification over the network’s structure composition.” (Sousa et al., p. 4) #5fb236
  * *
  
   
  
  - “This model is associated with a clArchitecture of a recurrent neural network composed of multiple clLayer instances, which are represented as clStructuralComponent elements.” (Sousa et al., p. 4) #5fb236
  * *
  
   
  
  - “The structural composition is further detailed using clEdge and clOrdering to capture the sequential or non-linear arrangement of components.” (Sousa et al., p. 4) #5fb236
  * *
  
   
  
  - “numerical prediction (e.g., surface area in hectares).” (Sousa et al., p. 4) #5fb236
  * *
  
   
  
  - “We define the training of the model with a clTrainingProcess, which relies on a clLossFunction (e.g., mean squared error) to guide optimization, and a clOptimizer (e.g., Adam) to update model parameters.” (Sousa et al., p. 4) #5fb236
  * *
  
   
  
  - “clHyperParameter” (Sousa et al., p. 4) #5fb236
  * *
  
   
  
  - “3) Prediction System metamodel - Dataset part:” (Sousa et al., p. 4) #2ea8e5
  * *
  
   
  
  - “Moreover, as data typically requires preprocessing and transformation before being used directly with neural networks, we support the definition of such transformations. Additionally, given that ecosystem data can be very limited for specific tasks, we also support generative transformations for data augmentation purposes.” (Sousa et al., p. 4) #5fb236
  * *
  
   
  
  - “annual surface area observations.” (Sousa et al., p. 4) #5fb236
  * *
  
   
  
  - “from high-level goals to a fully specified and generated software product.” (Sousa et al., p. 5) #5fb236
  * *
  
   
  
  - “Intent task” (Sousa et al., p. 5) #a28ae5
  * *
  
   
  
  - “Assisted Intent Recognition” (Sousa et al., p. 5) #5fb236
  * *
  
   
  
  - “Observer Elicitation” (Sousa et al., p. 5) #2ea8e5
  * *
  
   
  
  - “EcoSys model” (Sousa et al., p. 5) #5fb236
  * *
  
   
  
  - “This model serves as the semantic foundation that informs and constrains all subsequent decisions in the workflow.” (Sousa et al., p. 5) #5fb236
  * *
  
   
  
  - “Acquire and Prepare Data:” (Sousa et al., p. 5) #2ea8e5
  * *
  
   
  
  - “validating them against the requirements” (Sousa et al., p. 5) #ffd400
  *How are data validated against the requirements given in the EcoSys models? *
  
   
  
  - “Select and Configure AI” (Sousa et al., p. 5) #2ea8e5
  * *
  
   
  
  - “To support users who may not have their own datasets or AI models, the workflow is designed to be generic, modular, and interoperable. It may interface with external registries that provide curated and validated artifacts, such as model hubs or open data portals, enabling the integration of existing resources into the design process, even when user-specific artifacts are unavailable. The workflow triggers final system generation once all validated artifacts are available in a shared System State Context (SSC).” (Sousa et al., p. 5) #ffd400
  *It is necessary to clarify how this model will be used. Is there any running environment to execute the specified models? Are there any code generation steps? Who is the main user of the proposed approach? What are the roles that are involved in the show process? How are they intended to collaborate? *
  
   
  
  - “This structured approach prevents design mismatches, such as collecting incompatible data or selecting an unsuitable neural network model, by explicitly modeling dependencies between components.” (Sousa et al., p. 5) #5fb236
  * *
  
   
  
  - “Predict the yearly evolution of vegetation surface in Mersch Forest.” (Sousa et al., p. 5) #5fb236
  * *
  
   
  
  - “The ERPS metamodels (EcoSys and PredSys) are interdependent and their consistency must be managed all over the workflow.” (Sousa et al., p. 5) #5fb236
  * *
  
   
  
  - “This explicit representation prevents critical mismatches, such as defining prediction goals that remain unsupported even after preprocessing and augmentation, or selecting AI architectures misaligned with the data’s structure or semantics.” (Sousa et al., p. 5) #5fb236
  * *
  
   
  
  - “ResilForest” (Sousa et al., p. 7) #2ea8e5
  * *
  
   
  
  - “model introduces vegetation dryness as a key property.” (Sousa et al., p. 7) #5fb236
  * *
  
   
  
  - “ffective methodologies should support diverse modeling approaches, enabling not only goal-driven design but also data-driven and technologydriven exploration.” (Sousa et al., p. 7) #5fb236
  * *
  
   
  
  - “The workflow is designed to adapt and guide the user from their intent to a complete and consistent ERPS specification” (Sousa et al., p. 7) #5fb236
  * *
  
   
  
  - “Observer Elicitation” (Sousa et al., p. 7) #5fb236
  * *
  
   
  
  - “Model-Driven” (Sousa et al., p. 8) #ffd400
  * *
  
   
  
  - “”What kind of ecosystem resilience properties can I predict with this model?”” (Sousa et al., p. 8) #5fb236
  * *
  
   
  
  - “The Classify intent task within the Assisted Intent Recognition pool would classify the intent as to start the Select and Configure AI pool and the workflow execution would be:” (Sousa et al., p. 8) #ffd400
  *I understand, this needs to be abstract being a workshop paper. However, the paper stays disconnected with the real execution environments and technologies. It is not clear how the presented models and processes are actually used in practice further than the given abstract specifications. *
  
   
  
  - “2) Prediction System Modeling (PredSys)” (Sousa et al., p. 8) #ffd400
  *Where are the features of interests are specified? Where the prediction goals are given? *
  
   
  
  - “By structuring domain knowledge within the EcoSys metamodel, it provides a direct pathway for developing testable and robust predictive systems from established ecological principles.” (Sousa et al., p. 9) #a28ae5
  *That's interesting and I can agree with this. *
  
   
  
  - “from highlevel requirements to a fully specified system.” (Sousa et al., p. 10) #ffd400
  *I'm missing the link with the realy system that will execute what you have specified with the proposed approach. *