Agentic AI can guide climate-resilient retrofits under tight budgets

Researchers have developed an Agentic AI-based optimization framework designed to help public authorities and asset managers choose climate-resilient building retrofit strategies under budget constraints, as Europe faces mounting pressure to adapt existing buildings to heatwaves, heavy rainfall and other climate hazards.

Titled Agentic AI for Climate-Resilient Building Retrofit: A Multi-Hazard Optimization Framework and published in Technologies the study presents a policy-aligned, explainable artificial intelligence system that connects climate vulnerability assessment, multi-objective optimization and multi-agent decision support to guide building retrofit planning at portfolio scale.

AI system links climate policy to retrofit decisions

The research addresses a growing problem in the built environment: large numbers of existing buildings must be adapted to climate change, but owners, municipalities and public agencies often lack scalable tools to decide which retrofit actions should be prioritized first. The issue is especially urgent in Europe, where a large share of the current building stock is expected to remain in use by 2050.

Climate adaptation, as the authors argue, is no longer only a question of identifying hazards, but a question of making timely, affordable and technically defensible decisions across large portfolios of buildings. Heatwaves can damage materials, increase cooling demand and worsen indoor comfort. Heavy precipitation can expose weaknesses in roofs, facades, openings, drainage systems and moisture protection. Yet retrofit decisions are often made with limited budgets, fragmented data and conventional tools that do not fully account for trade-offs.

The framework developed in the study is aligned with European Union guidance on climate adaptation for buildings. It begins with a Climate Vulnerability Assessment that evaluates building exposure, sensitivity and adaptive capacity. These factors are then translated into a vulnerability index that can be used in an optimization model. The framework is designed to identify retrofit strategies that reduce vulnerability while keeping costs within a defined budget.

The study focuses on two climate hazards, heatwaves and heavy precipitation, because they create different and sometimes conflicting retrofit needs. A measure that helps one hazard may worsen another. For example, some insulation choices can improve heat protection but create moisture-related concerns under heavy rainfall conditions. This makes climate-resilient retrofit planning a multi-hazard problem rather than a simple checklist of improvements.

The proposed system evaluates building envelope interventions including green roofs, blue roofs, roof insulation, window replacement, shading systems, hail-proof systems, external thermal insulation and water-repellent walls. The study treats each retrofit option as a discrete decision, meaning an action is either selected or not selected for a building. That makes the problem computationally complex as the number of buildings and possible retrofit combinations grows.

To test the approach, the researchers generated synthetic data for 50 buildings, using vulnerability ranges based on expert-informed assessment rules. They then tested retrofit optimization under an EUR 800,000 budget with eight possible retrofit actions. The use of synthetic data allowed the authors to stress-test the framework before real-world deployment, but it also remains one of the study’s main limitations.

MO-IWO outperforms faster Grey Wolf model

The study compares two multi-objective metaheuristic algorithms: Multi-Objective Invasive Weed Optimization and Multi-Objective Grey Wolf Optimization. Both were adapted to handle the dual goal of minimizing total building vulnerability and minimizing retrofit cost.

The results show a clear trade-off between speed and quality. The Grey Wolf model required less computation time, completing the test in 91.27 seconds, compared with 363.30 seconds for the Invasive Weed model. However, the Invasive Weed model produced stronger optimization results across almost every performance measure.

The researchers found that Multi-Objective Invasive Weed Optimization achieved a higher-quality Pareto front, the set of possible trade-off solutions between cost and risk reduction. It delivered about 31.5 percent higher hypervolume, a measure of how well the optimizer covers the solution space, and 65.8 percent lower inverted generational distance, which indicates closer and more uniform convergence toward the reference front. It also posted stronger system stability, at 94.74 percent, compared with 83.94 percent for the Grey Wolf model.

The outcomes led the authors to select the Invasive Weed model as the mathematical optimization engine for the Agentic AI system. The finding matters because retrofit planning cannot be judged only by speed. For public authorities and infrastructure managers, a faster model may be less valuable if it produces weaker or less stable decision options.

The framework also tested several methods for choosing a final solution from the Pareto front, including max distance, utopia, weighted sum, TOPSIS, VIKOR, marginal gain and angle-based selection. These methods reflect different decision priorities. Some favor balanced cost-risk trade-offs, while others prioritize lower vulnerability at higher cost or lower spending with more residual risk.

In the standalone optimization, several decision methods converged on a common compromise solution under the Invasive Weed model, suggesting a robust knee point in the cost-vulnerability trade-off. TOPSIS tended to select lower-vulnerability, higher-cost solutions. The angle method leaned toward lower-cost options with higher remaining vulnerability. Marginal gain identified points where additional spending produced the strongest return in vulnerability reduction.

The study’s sensitivity analysis found that external thermal insulation emerged as one of the most influential retrofit actions in the tested scenario. The results indicated that the framework’s recommended strategy remained stable even under modest changes in cost or technical effectiveness. That stability is important for real-world use, where material prices, labor costs and performance assumptions can change.

Agentic AI turns optimization into decision support

What sets the study apart is its multi-agent layer surrounding the optimizer. Rather than presenting users with a static model that requires technical inputs, the system uses specialized AI agents to interpret user goals and convert them into structured optimization parameters.

The architecture includes an Orchestrator, a Requirements Agent, a Cost Agent, a Strategy Agent and an XAI Agent.

• The Requirements Agent identifies whether the user wants analysis for a single building, a subset of buildings or an entire portfolio.
• The Cost Agent extracts budget limits and standardizes financial values.
• The Strategy Agent selects the decision method, determines whether a quick or deep analysis is needed, and translates qualitative user priorities into numerical weights.
• The XAI Agent converts numerical results into human-readable explanations and can generate structured reports.
• The Orchestrator coordinates these agents and merges their outputs into a global configuration for the optimization engine. The system is designed to handle incomplete or iterative natural-language queries.

This design allows users to interact with the system through natural language queries. A decision-maker could ask for a quick optimization for specific buildings, a deep analysis across the full portfolio, a revised budget scenario, or a comparison of decision methods. The framework can also reuse previous results when only the final decision method changes, reducing the need for repeated computation.

The study focuses on human control. If a user requests a change to input data, such as modifying the cost of a retrofit action, the system holds the change in a pending state until the user confirms it. After confirmation, the system creates a new version of the input dataset while preserving the original. This audit-trail structure is intended to reduce black-box risk and ensure that decisions remain traceable.

The Agentic AI layer changed the behavior of the optimization process. Traditional optimization tends to use most of the available budget in pursuit of the lowest possible vulnerability. The Agentic AI system behaved more like a decision-support consultant, stabilizing around roughly 50 percent budget utilization in the tested scenario. According to the authors, this suggests that the system identified a point of diminishing returns, where further spending produced only limited additional vulnerability reduction.

Public agencies often operate under budget uncertainty and competing social priorities. A system that treats the budget as a maximum ceiling, not a target to exhaust, may help preserve resources for other assets or later interventions. At the same time, the study notes that such conservative behavior may not always align with political or social goals where minimizing vulnerability is the top priority.

It is important to note that the framework currently relies on static cost databases, which may not reflect market volatility. The vulnerability data used in the proof of concept were stochastically generated rather than drawn from a real building portfolio. The hazard exposure component was simplified to preserve computational scalability. Cross-hazard interactions also require more experimental evidence, especially where one retrofit action may reduce one risk while increasing another.

Future research will focus on integrating the system with Building Information Modeling to automate data extraction, connect building quantities with government price databases and improve real-world cost accuracy. The authors also plan to move from a vulnerability index to a broader risk index that accounts for social consequences, such as population density, strategic importance of buildings and service disruption. That would allow hospitals, schools and warehouses to be ranked differently based not only on physical exposure but also on the consequences of failure.

The study also points to the future use of Monte Carlo simulations or fuzzy logic to quantify uncertainty, along with life-cycle cost analysis using net present value, internal rate of return and payback period. These additions would help decision-makers judge not only which retrofit actions reduce climate risk, but whether those investments remain economically sustainable over the full life of a building.