Regulatory-Grade PFAS Lab Automation with an Analytical Decision Layer - Exception-Based Review for EPA 1633 and EPA 533 Using Limited Sample Models and an AI Co-Pilot Workflow

AI to Increase Lab Productivity in Environmental Testing
Oral Presentation

Prepared by L. Theverapperuma
Expert Intelligence, 2445 Augustine Dr. Suite 150, Santa Clara, California, 95051, United States


Contact Information: [email protected]; 408-781-2656

ABSTRACT

Environmental laboratories face rising PFAS testing volume, fixed staffing, and persistent variability in chromatographic integration and data review. While large spectral libraries and large-data AI systems can reduce analyst burden, most laboratories operate with limited labeled data and incomplete reference coverage, especially for complex cases involving coelution, qualifier ion ratio distortion, retention time drift, and branched isomer behavior. This creates operational risk - rework, delayed turnaround time, and missed detections for emerging PFAS.

We present a practical, trust-first automation pattern for PFAS LC-MS/MS that separates (1) signal interpretation from (2) workflow and documentation. A Limited Sample Model (LSM) serves as the analytical decision layer - converting raw chromatograms and transitions into evidence artifacts including peak boundaries, peak quality scores, interference flags, retention time checks, and quantifier-qualifier consistency metrics. The decision layer applies method-specific QC gates aligned to EPA 1633 and EPA 533 - calibration and QC acceptance, internal standard recovery checks, retention time windows, qualifier ion ratio tolerances, blanks and carryover checks - and routes only exceptions for analyst review.

An AI co-pilot layer then supports productivity outside the measurement engine - triaging the review queue, generating evidence-grounded explanations, and drafting audit-ready notes using only computed decision artifacts, while final analytical decisions remain with the human analyst.

Results from routine PFAS batches show high agreement with expert review for qualitative identification and peak decisioning (F1-score 0.994), with meaningful reduction in manual review workload via exception-based routing. Training and evaluation were performed using modern GPU infrastructure (NVIDIA H100-class), enabling fast iteration while maintaining versioned model governance and traceable evidence bundles.

This talk focuses on real-world deployment considerations - validation strategy, failure modes, and how an analytical decision layer can increase throughput and consistency without compromising defensibility or public trust.

References
EPA Method 1633A - PFAS in aqueous, solid, biosolids, and tissue by LC-MS/MS.
EPA Method 533 - PFAS in drinking water by SPE LC-MS/MS.