Differentiable mass transfer modeling enables wet lab validated inverse analysis of liquid chromatography
Year: 2026
Type: preprint
Abstract: Liquid chromatography (LC) is essential to organic synthesis and emerging selfdriving laboratories (SDLs), yet LC method development remains cost-intensive and dependent on expert intuition. Mechanistic chromatographic models offer a route to transferable and data-efficient method development, but their practical use is limited by the computational cost of inverse modeling and by the ill-posedness of parameter estimation from experimental chromatograms. Here we report a wet-lab-validated differentiable physics framework for LC inverse modeling. The framework integrates mechanistic column modeling with physics-informed machine learning (PIML) to enable efficient forward simulation and gradient-based inverse estimation of molecular transport and adsorption parameters from gradient-elution chromatograms. The primary PINN surrogate reproduces the mechanistic reference with a root-mean-square error of 10−3 in dimensionless concentration. To address the ill-posed nature of the inverse problem, we introduce an observability-guided experimental-design strategy that quantifies parameter identifiability and incorporates this information into the inverse objective to promote convergence toward the ground-truth parameters. Across 300 dry-lab simulation parameter combinations, the proposed strategy recovered the ground-truth parameters with relative errors ≤ 10% in 88.3% of cases using three gradient programs, compared with only 6.7% using a single gradient program. We further validate the framework against wet-lab LC 1 measurements, achieving an absolute retention-time error ≤ 10 s for 89.3% of samples and establishing a wet-lab database comprising 731 sets of transport and adsorption parameters extracted from 2,553 gradient-elution LC chromatograms. This work provides an observability-aware differentiable modeling foundation for autonomous LC method development and closed-loop analytical decision-making in AI-driven laboratories.
Source: ChemRxiv
Institution: Chinese University of Hong Kong
Language: English
Subfield: Analytical Chemistry
Field: Chemistry
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7214085388