ADS-net: Accelerated Physisorption Prediction via Deep Learning

ADS-net* enables the rapid calculation of adsorption maps for the prediction of physisorption isotherms. It is based on a morphological adsorption model** that simulates gas molecule adhesion by dilation operations (adsorbed multilayer film) and gas condensation by critical pore size detection by closing operations. Connected component labelling is used to model percolation and hysteresis.

The network architecture was optimised for this task with subpixel depth upscaling layers and a perceptual loss function. In addition, 3D geodesic information was added to the training dataset prior to slicing to improve the performance of the 2D model for 3D prediction.

Physisorption isotherms are obtained by applying the inverse Kelvin-Cohan equation to the predicted adsorption maps. This method allows large 3D microstructures to be simulated with high efficiency, reducing computational time by more than 30x compared to the standard model** for a 512³ voxel volume.

Three plug-ins are provided:
ADS-net Computation Module (ADS-net CPU): generates adsorption maps from 3D microstructures. This is a CPU version inference with ONNX framework.
Distance Augmentation Module (Distance augmentation): adds distance field information to an existing binary 3D microstructure.
Physisorption Curve Construction Module (Physisorption from ADS-net): calculates isotherms from ADS-net adsorption maps.

ADS-net was trained on Boolean model microstructures at a resolution of 0.36 nm per voxel. Further details can be found in *. Inference can be performed either patch-wise (for higher quality), as suggested in the paper, or by a single inference per plane. The full training procedure and all mandatory files are provided with the link below.

A sample volume is also provided for testing the module, with distance map enrichment already implemented.

* A Hammoumi, M Moreaud, E Jolimaitre, T Chevalier, M Klotz, A Novikov. Accelerating a Morphology-Preserving Adsorption Model by Deep Learning. 2022 IEEE International Conference on Image Processing (ICIP), 1851-1855 (2022).

** A. Hammoumi, M. Moreaud, D. Jeulin, E. Jolimaitre, T. Chevalier, L. Sorbier, M. Klotz, A. Novikov. A novel physisorption model based on mathematical morphology operators preserving exact pore morphology and connectivity. Microporous and Mesoporous Materials (2022).