Building upon this broad perspective on sorbents, our study focuses specifically on desiccants. In response to the limitations of traditional methods, we introduce a novel one-shot experimental approach that employs a single and continuous adsorption-desorption protocol over a partial pressure range normalized by the saturation pressure from 0 to 0.9. This method captures broad transient mass uptake data in a single experimental run, enabling the interpolation of the complete equilibrium isotherm without resorting to the laborious stepwise procedure. The susceptibility of desiccants to performance degradation over time makes this simple, one-shot experimental procedure a promising new method, especially given the substantial recent increase in research interest for desiccants, thanks to their applications in atmospheric water harvesting, thermal energy storage, and cooling. The experimental data is obtained using high-resolution sorption systems, a vacuum dynamic vapor sorption (DVS), and multi-species advanced sorption analyzers to reduce model error and isolate the impact of mixed species. The model prediction is then compared to the isotherms gained from experimental data in the literature, based on traditional multi-step methods, to demonstrate that our one-shot approach is not only more efficient but also reliably reconstructs the full adsorption isotherm for desiccants.
To achieve this goal, we have developed a dual-framework strategy that integrates inverse modeling and physics-informed neural networks (PINNs). The inverse modeling approach redefines the sorption process as an optimization problem, wherein equilibrium isotherm parameters are iteratively tuned to minimize discrepancies between the observed transient uptake and model predictions. This approach extracts the equilibrium isotherm of desiccants from a single one-shot experiment by exploiting the non-steady time-series dynamic information present in the transient data, eliminating time-consuming and laborious efforts to ensure equilibrium at each partial pressure as required in the traditional analytical approach. In other words, while traditional approaches wait for a steady state to extract a single value from a single experiment, we propose inferring model frameworks from the information contained in the entire time-series profile. This approach may also resolve the critical issue of in-situ material degradation estimation that can occur in a long operating condition. Complementing the inverse approach, we further developed a PINN framework to analyze complex physical behaviors, including isotherm-type characterization and mixed species interaction. PINN embeds the governing equations of mass transport and adsorption kinetics directly into the neural network architecture, improving accuracy and robustness. We implemented a recently proposed novel architecture of Kolmogorov–Arnold Networks (KAN) to derive symbolic expressions, leading to better interpretation than regular neural networks. This physics-informed strategy ensures that the trained model remains faithful to the underlying physical principles, thereby improving the predictive accuracy of the isotherm reconstruction even under variable experimental conditions.
Comparative analyses reveal that this method not only reduces the experimental duration and resource consumption but also provides insights into the interplay between kinetic effects and equilibrium behavior and opens the prospect for implementation for material performance estimation. As a result, the proposed framework offers a powerful tool for rapid screening and optimization of desiccant materials, potentially reducing up to 90% of characterization time, and ultimately contributing to improved performance in gas separation, direct air capture, thermal energy storage and management, and atmospheric water harvesting.