Transformer-based prediction of temperature-dependent adsorption equilibrium on porous adsorbents in semiconductor cleanroom environments

This study comparatively analysed the adsorption mechanisms of ammonia, SO 2 , and water on M-2 in semiconductor cleanroom environments. Results show that ammonia's experimental/modelled adsorption isotherms at 288/298/308 K are optimally fitted by the Dual-Site Langmuir isotherm model (DSLIM), with corresponding root mean square error (RMSE) values of 0.109, 0.132 and 0.063. DSLIM also best fits SO 2 adsorption at 288 K/298 K (RMSE: 0.076/0.058), while the Toth isotherm model performs optimally at 308 K (RMSE: 0.028). Calculations reveal ammonia's initial adsorption heat is 89.80 kJ/mol, indicating low-capacity chemisorption; SO 2 's initial adsorption heat is below 40 kJ/mol, while the adsorption heat of water remains consistently close to its condensation heat (45 kJ/mol). It should be noted that this observation is based on the isosteric heat calculated from isotherm fitting, which inherently yields an averaged value of multiple adsorption interactions. Based on this averaged result, it can be inferred that water exerts a weak yet persistent competitive effect during the adsorption process. Thermodynamic analysis shows pre-adsorbed water cannot displace adsorbed ammonia. The transformer-based sequence model predicts adsorption capacity with mean absolute percentage error of 15.47%–29.13% for ammonia and 4.82%–10.45% for SO 2 , showing good predictability and transferability, supporting its application in semiconductor cleanroom adsorption system optimization.