Search the EPA Research Database

Project title

F-gases and Chemicals Human-in-the-loop Artificial Intelligence Surveillance System Design

Primary Funding Agency

Environmental Protection Agency (EPA)

Co-Funding Organisation(s)

n/a

Lead Organisation

South East Technological University (SETU)

Lead Applicant

Vibhutesh Kumar Singh

Project Abstract

Illegal import of fluorinated gases (F-Gases) and other chemicals that fall under the F-Gas, REACH, and POP Regulations pose environmental and public health risks. Enforcement projects from the European Chemicals Agency (ECHA) strongly suggest that the regulations are not being followed, particularly in the products ordered from online marketplaces. These imports often enter European and Irish markets through indirect supply chain routes facilitated by vague product labeling and can have complex chemical blends and mixtures. Current manual enforcement methods, which include physical audits, laboratory testing, and product supply chain inspections, are slow and insufficient to monitor these illegal sales effectively. We propose the F-gases and Chemicals Human-in-the-loop (HITL) Artificial Intelligence Surveillance System (F-CASS) to address the challenges of restricted chemical surveillance. F-CASS will use a multimodal large language model (LLM) agent that can autonomously invoke open-source intelligence (OSINT) tools for targeted data acquisition of the product under suspicion and similar product listings in a General Data Protection Regulation (GDPR) compliant manner. The acquired data, along with a chemical domain-specific knowledge graph and regulatory heuristics, will be used to calculate risk and flag potentially illegal products. These regulatory heuristics consider various indicators, e.g., marketplace history, country of origin, suspicious shipping routes, pricing anomalies, inconsistent manufacturer information, misleading product descriptions or images, and even suspicious customer reviews. Customized OSINT web crawlers and scrapers will form the frontline of this system to periodically or on-demand scan a wide range of digital platforms for illegal chemical trade, including regular e-commerce websites, chemical sales portals, and the product manufacturer's website. The HITL design of this proposal is a unique aspect, as LLMs are usually trained with a wide range of data, they may lack expert knowledge of the chemical regulatory domain; our approach includes chemical and regulatory domain expert insights to establish critical points of detection, particularly chemistry knowledge graphs for complex compound mixtures and regulatory heuristics. Reasoning LLMs will be used in the F-CASS system to explain how they reached a decision, which will help the end-users connect the dots and understand which parts of the knowledge graph and regulatory heuristic were triggered. The HITL aspect of the project will support the trustworthiness of our system in the regulatory domain and help remove the 'black box' stigma of AI systems. This proposal employs engaged research methods, such as co-design workshops with authorities like the EPA, to enable policymakers to use the project's key outputs directly. These outputs will include a prototype AI toolkit for surveillance of restricted chemicals, a review of best practices in market surveillance, and recommendations to support policymakers in improving national and EU-level strategies to develop and implement coherent policy approaches to control the illegal imports of these restricted chemicals.

Grant Approved

€161,493.51

Research Hub

Delivering a Healthy Environment

Research Theme

n/a

Start Date

01/03/2026

Initial Projected Completion Date

31/08/2027