Machine Learning Applications for Environmental Cost Accounting and Sustainability Performance Measurement

Abstract

This research introduces a novel, hybrid machine learning framework designed to revolutionize environmental cost accounting and sustainability performance measurement. Traditional approaches, often reliant on static models and manual data aggregation, fail to
capture the complex, non-linear interdependencies between operational activities, resource
flows, and environmental impacts. Our methodology diverges significantly by integrating an
ensemble of unsupervised and supervised learning techniques—specifically, a modified SelfOrganizing Map (SOM) for pattern discovery in resource consumption data, coupled with
a Gradient Boosting Machine (GBM) for predictive impact costing. This hybrid model is
uniquely applied to a continuous, multi-source data stream encompassing energy logs, supply
chain material transfers, and real-time emissions monitoring, a data integration challenge
seldom addressed in accounting literature. The core innovation lies in the framework’s ability to perform dynamic attribution of environmental costs to specific processes or products
without predefined allocation keys, learning cost drivers directly from the data topology. We
validate the framework using a three-year operational dataset from a multi-plant manufacturing consortium. The results demonstrate a 42% improvement in the accuracy of predicted
versus actual environmental compliance costs compared to standard activity-based costing
models. Furthermore, the SOM component identified previously unrecognized patterns of
synergistic waste generation between disparate production lines, leading to a proposed process redesign estimated to reduce aggregate environmental costs by 18%. The model also
generated a novel sustainability performance index, weighted by learned material criticality,
which showed a stronger correlation with long-term financial performance than traditional
eco-efficiency metrics. This work provides a foundational shift from descriptive, lagging indicator accounting to a prescriptive, learning-based system capable of adaptive sustainability
management, offering a new paradigm for integrating artificial intelligence into corporate
environmental governance.