Accelerating scope 3 emissions accounting: LLMs to the rescue

The rising interest in the calculation and disclosure of Scope 3 GHG emissions has thrown the spotlight on emissions calculation methods. One of the more common Scope 3 calculation methodologies that organizations use is the spend-based method, which can be time-consuming and resource intensive to implement. This article explores an innovative way to streamline the estimation of Scope 3 GHG emissions leveraging AI and Large Language Models (LLMs) to help categorize financial transaction data to align with spend-based emissions factors.

Scope 3 emissions, also called indirect emissions, encompass greenhouse gas emissions (GHG) that occur in an organization’s value chain and as such, are not under its direct operational control or ownership. In simpler terms, these emissions arise from external sources, such as emissions associated with suppliers and customers and are beyond the company’s core operations.

A 2022 CDP study found that for companies that report to CDP, emissions occurring in their supply chain represent an average of 11.4x more emissions than their operational emissions.

The same study showed that 72% of CDP-responding companies reported only their operational emissions (Scope 1 and/or 2). Some companies attempt to estimate Scope 3 emissions by collecting data from suppliers and manually categorizing data, but progress is hindered by challenges such as large supplier base, depth of supply chains, complex data collection processes and substantial resource requirements.

One approach to estimating Scope 3 emissions is to leverage financial transaction data (for example, spend) as a proxy for emissions associated with goods and/or services purchased. Converting this financial data into GHG emissions inventory requires information on the GHG emissions impact of the product or service purchased.

The US Environmentally-Extended Input-Output (USEEIO) is a lifecycle assessment (LCA) framework that traces economic and environmental flows of goods and services within the United States. USEEIO offers a comprehensive dataset and methodology that merges economic IO analysis with environmental data to estimate the environmental consequences associated with economic activities. Within USEEIO, goods and services are categorized into 66 spend categories, referred to as commodity classes, based on their common environmental characteristics. These commodity classes are associated with emission factors used to estimate environmental impacts using expenditure data.

The Eora MRIO (Multi-region input-output) dataset is a globally recognized spend-based emission factor set that documents the inter-sectoral transfers amongst 15.909 sectors across 190 countries. The Eora factor set has been modified to align with the USEEIO categorization of 66 summary classifications per country. This involves mapping the 15.909 sectors found across the Eora26 categories and more detailed national sector classifications to the USEEIO 66 spend categories.

This is where LLMs come into play. In recent years, remarkable strides have been achieved in crafting extensive foundation language models for natural language processing (NLP). These innovations have showcased strong performance in comparison to conventional machine learning (ML) models, particularly in scenarios where labelled data is in short supply. Capitalizing on the capabilities of these large pre-trained NLP models, combined with domain adaptation techniques that make efficient use of limited data, presents significant potential for tackling the challenge associated with accounting for Scope 3 environmental impact.

Our approach involves fine-tuning foundation models to recognize Environmentally-Extended Input-Output (EEIO) commodity classes of purchase orders or ledger entries which are written in natural language. Subsequently, we calculate emissions associated with the spend using EEIO emission factors (emissions per $ spent) sourced from Supply Chain GHG Emission Factors for US Commodities and Industries for US-centric datasets, and the Eora MRIO (Multi-region input-output) for global datasets. This framework helps streamline and simplify the process for businesses to calculate Scope 3 emissions.

Figure 1 illustrates the framework for Scope 3 emission estimation employing a large language model. This framework comprises four distinct modules: data preparation, domain adaptation, classification and emission computation.

We conducted extensive experiments involving several cutting-edge LLMs including roberta-base, bert-base-uncased, and distilroberta-base-climate-f. Additionally, we explored non-foundation classical models based on TF-IDF and Word2Vec vectorization approaches. Our objective was to assess the potential of foundation models (FM) in estimating Scope 3 emissions using financial transaction records as a proxy for goods and services. The experimental results indicate that fine-tuned LLMs exhibit significant improvements over the zero-shot classification approach. Furthermore, they outperformed classical text mining techniques like TF-IDF and Word2Vec, delivering performance on par with domain-expert classification.

Employing LLMs in the process of estimating Scope 3 emissions is a promising new approach..

As previously explained, spend data is more readily available in an organization and is a common proxy of quantity of goods/services. However, challenges such as commodity recognition and mapping can seem hard to address. Why?

• Firstly, because purchased products and services are described in natural languages in various forms, which is why commodity recognition from purchase orders/ledger entry is extremely hard.
• Secondly, because there are millions of products and service for which spend based emission factor may not be available. This makes the manual mapping of the commodity/service to product/service category extremely hard, if not impossible.

Here’s where deep learning-based foundation models for NLP can be efficient across a broad range of NLP classification tasks when availability of labelled data is insufficient or limited. Leveraging large pre-trained NLP models with domain adaptation with limited data has potential to support Scope 3 emissions calculation.

In conclusion, calculating Scope 3 emissions with the support of LLMs represents a significant advancement in data management for sustainability. The promising outcomes from employing advanced LLMs highlight their potential to accelerate GHG footprint assessments. Practical integration into software like the IBM Envizi ESG Suite can simplify the process while increasing the speed to insight.