⚡ electrical engineering

Explainable Artificial Intelligence Framework for Predicting Construction Waste and Embodied Carbon in Building Projects Using Literature Calibrated Simulation

This study presents an explainable artificial intelligence framework that utilizes a literature-calibrated simulation of 1,200 building archetypes to accurately predict construction waste and embodied carbon intensity through advanced regression algorithms and SHAP-based feature analysis, demonstrating how integrating material data enhances predictive accuracy for early-stage decision support.

Original authors: Amitkumar D. Raval¹, Jagruti Shah, Arti N. Pamnani

Published 2026-09-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Amitkumar D. Raval¹, Jagruti Shah, Arti N. Pamnani

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Every time a building rises from the ground, it leaves a trail of two distinct but deeply connected footprints. One is the pile of leftover bricks, cut-off pipes, and packaging that ends up in a landfill; the other is the invisible cloud of greenhouse gases released to make the steel, concrete, and glass that form the walls. For decades, engineers and architects have treated these two outcomes as separate problems. They might calculate how much waste a project will generate while using a completely different set of tools to estimate its carbon emissions. Yet, the choices that create waste—ordering too much material, cutting it poorly, or changing the design late in the game—are the very same choices that determine how much pollution is locked into the structure. To truly build for the future, the industry needs a way to see both footprints at the same time, right from the earliest sketches, before the first shovel hits the dirt.

A team of researchers from India has developed a new way to do exactly that. They created a digital framework that uses artificial intelligence to predict both construction waste and the carbon emissions of a building's materials simultaneously. The challenge they faced was a lack of real-world data. In the messy reality of construction sites, records of exactly how much waste was generated and the precise carbon cost of every single material are rarely kept in a way that can be easily compared. Because no public database existed with this specific, aligned information, the researchers did not try to guess or fabricate records. Instead, they built a transparent, computer-generated simulation. They created 1,200 virtual building projects, ranging from small offices to large hospitals, and assigned them realistic characteristics like structural systems, climate zones, and construction methods. For each virtual project, they used established scientific formulas to calculate exactly how much waste and carbon would be produced, creating a perfect "ground truth" that they could then test their predictions against.

The researchers trained six different types of machine-learning algorithms to look at the basic details of a project—such as its size, the number of floors, the type of structure, and the local climate—and guess the waste and carbon outcomes. They tested these algorithms in two different scenarios. In the first scenario, the computer only knew the early-stage design details, similar to what a client knows when they first hire an architect. In the second scenario, the computer was given the added benefit of knowing the specific amounts of materials, like the exact weight of steel or concrete required. The results showed that even with just the early design details, the artificial intelligence could make surprisingly accurate predictions. For construction waste, the best model could predict the amount of trash with an error margin of less than 12 percent. However, predicting the carbon emissions was harder without knowing the specific material quantities. When the computer was given the extra information about material amounts, its ability to predict carbon emissions improved dramatically, cutting the error rate in half.

What makes this work particularly valuable is that the researchers did not just build a "black box" that spits out a number. They equipped their system with a feature that explains why it made a specific prediction. By analyzing the data, the system identified the most important drivers behind the waste and carbon numbers. It found that the type of structural system—whether the building is made of reinforced concrete, steel, or wood—was the single biggest factor. The number of floors, the amount of recycled material used, and the ratio of windows to wall area also played major roles. For instance, the simulation confirmed that using more prefabricated parts, which are made in controlled factories, generally leads to less waste on the construction site. These explanations allow architects and engineers to understand the levers they can pull. They can see that choosing a different structural system or increasing the use of recycled content will directly change the predicted outcomes.

The study also revealed a crucial nuance about how these tools should be used. The researchers found that the most complex, powerful algorithms were not always the best choice. For predicting carbon emissions, a simpler, more straightforward mathematical model performed just as well as the complex ones once the material quantities were known. This suggests that in the real world, the best tool depends on the information available. If a team is in the very early stages of design, a sophisticated model that can find hidden patterns in general project details is useful. But once the material quantities are known, a simpler, transparent model might be just as accurate and easier to trust. The researchers emphasize that their results are based on a controlled simulation, not on a specific real-world city or market, so the exact numbers should not be taken as a guarantee for every future building. However, the framework itself proves that it is possible to combine waste and carbon prediction into a single, explainable system. This approach offers a path forward for the construction industry to make smarter, more sustainable decisions earlier in the process, ensuring that the buildings of tomorrow are not only efficient but also cleaner and less wasteful from the very first line drawn on a blueprint.

Technical Summary: Explainable AI Framework for Predicting Construction Waste and Embodied Carbon

Problem Statement
Construction waste and embodied carbon are critical sustainability metrics in the building sector, yet they are typically estimated through separate workflows despite arising from interdependent design and procurement choices. Current research faces four primary gaps: (1) waste and carbon are rarely modeled jointly, limiting insight into shared drivers; (2) high-performing algorithms often function as "black boxes," lacking the explainability required for design teams; (3) performance comparisons are hindered by inconsistent data maturities (ranging from concept descriptors to detailed material quantities); and (4) a scarcity of aligned, open project-level datasets containing both waste and cradle-to-practical-completion (modules A1–A5) carbon observations. The lack of empirical data often leads to the temptation of fabricating records or overclaiming results from small, opaque datasets.

Methodology
To address the data gap without fabricating empirical records, the authors conducted a controlled, literature-calibrated simulation.

Data Generation: A dataset of 1,200 building-project archetypes was generated using a fixed random seed. The dataset spans five building uses, three structural systems, three procurement routes, and four climate zones.
Target Generation:
- Construction Waste: Calculated using explicit material-flow equations modified by management factors (prefabrication, BIM maturity, waste planning) and stochastic uncertainty.
- Embodied Carbon (A1–A5): Calculated using generic emission factors for materials, transport, and construction, adjusted for recycled content and logistics.
Scenarios: Two information scenarios were tested to avoid information leakage:
- Scenario S1 (Concept-stage): Uses project descriptors (e.g., use, structure, floor count, climate, BIM maturity) available early in design.
- Scenario S2 (Material-aware): Adds seven specific material intensities (kg/m²) available at schematic design or via BIM.
Modeling: Six regression algorithms were evaluated: Linear Regression, Elastic Net, Random Forest, Extremely Randomized Trees, XGBoost, and LightGBM. Models were trained on an 80/20 hold-out split with three-fold cross-validation.
Explainability: SHapley Additive exPlanations (SHAP) were used to identify global feature importance and local predictive drivers.

Key Results

Waste Prediction: In the concept-stage scenario, XGBoost achieved the highest performance ( $R^2 = 0.870$ , MAPE = 11.89%). Adding material intensities (S2) yielded only a modest improvement ( $R^2$ increased to 0.888), suggesting that concept-stage descriptors (structural system, floor count, prefabrication) already capture most of the variance in waste generation.
Embodied Carbon Prediction: Concept-stage prediction was more challenging, with Linear Regression achieving the highest $R^2$ (0.694) and LightGBM the lowest MAE (21.73 kgCO₂e/m²). However, adding material intensities (S2) significantly improved accuracy, with Elastic Net reaching an $R^2$ of 0.906 and MAPE dropping to 4.27%. This indicates that embodied carbon is highly dependent on specific material quantities, which are often unavailable at the earliest design stages.
Explainability: SHAP analysis identified structural system, floor count, recycled content, facade-to-floor ratio, and prefabrication as dominant predictors. In the material-aware scenario, specific material intensities (steel, concrete, glass) became the primary drivers for carbon predictions. The study notes that while structural system is a strong proxy, it does not imply a causal preference for one system over another without further context.

Key Contributions
The paper makes four specific methodological contributions:

Unified Schema: A data schema that aligns construction waste and embodied carbon for joint prediction.
Transparent Simulation: A reproducible, literature-calibrated simulation framework that avoids presenting invented observations as empirical evidence, explicitly labeling the data as synthetic.
Benchmarking: A comparative analysis of linear, regularized, and tree-ensemble models across different information maturities (concept vs. material-aware).
Explainability Layer: The integration of SHAP to translate model predictions into decision-relevant drivers, distinguishing between predictive associations and causal interventions.

Significance and Claims
The authors position this work as a methodological template rather than a deployable estimator for real-world projects. The significance lies in demonstrating how predictive accuracy, information maturity, and transparent explanations can be combined in a unified decision-support architecture.

Staged Information: The study validates that early-stage predictions are feasible but inherently limited by missing material data; accuracy improves significantly as design matures.
Algorithm Selection: The results challenge the assumption that complex nonlinear models are always superior. For embodied carbon with known material quantities, additive linear models (Elastic Net) performed competitively, suggesting algorithm choice should be driven by data availability and decision requirements rather than novelty.
Governance: The framework supports governance by exposing model behavior, allowing for plausibility checks, and identifying data gaps. It emphasizes that explanations should support professional judgment rather than replace it.

The authors explicitly state that the reported error values should not be cited as expected performance on real projects due to the synthetic nature of the data, the exclusion of product-specific Environmental Product Declarations (EPDs), and the lack of regional calibration. The next necessary step identified is validation using measured waste records and harmonized carbon assessments across multiple regions.

Technical Summary: Explainable AI Framework for Predicting Construction Waste and Embodied Carbon

More like this