Speaker
Description
Traditional methods of characterization of nanoporous carbons are based on the use of a simulated kernel of isotherms obtained by classical density functional theory (cDFT) in a series of independent slit or cylindrical pore models to compute the pore size distribution. This approach cannot describe the structure's asymmetry, a feature characteristic of these materials. Vallejos et al. [1] used 3D carbon structure models to generate kernels in recent work. They found relative contributions of these structures in experimental isotherms, calculating morphological parameters of the experimental sample and plausible simulated (Transmission Electron Microscopy) TEM images. In this work, we try to predict molecular structures of real materials from a set of experimental data (adsorption properties and images from TEM) using score-based diffusion models. The score-based diffusion model is a deep generative model that has achieved state-of-the-art sample quality in several tasks, including image generation by Computed Tomography (CT) and Magnetic Resonance Imaging (MRI). In this approach, we perturb data with a sequence of noise distributions and generate samples by learning to reverse this path from noise to data. The framework of score-based diffusion models involves gradually diffusing the data distribution towards a given noise distribution using a stochastic differential equation (SDE) and learning the time reversal of this SDE for sample generation. Crucially, the reverse-time SDE has a closed-form expression that depends solely on a time-dependent gradient field, called the score, of the perturbed data distribution. This gradient field can be efficiently estimated by training a neural network from a score-based model with a weighted combination of score-matching losses as the objective. The coupling of adsorption properties and images from TEM with score-based generative models to solve the inverse problems related to the reconstruction of the porous material structure has enormous potential for broadly impacting the porous material characterization area.
References:
1. F. Vallejos-Burgos, C. de Tomas, N. J. Corrente, K. Urita, S. Wang, C. Urita, I. Moriguchi et al. 3D nanostructure prediction of porous carbons via gas adsorption. Carbon, 215 (2023) 118431
2. Y. Song, C. Durkan, I. Murray, S. Ermon. Maximum likelihood training of score-based diffusion models. Advances in Neural Information Processing Systems, 34 (2021) 1415-1428
3. Y. Song, L. Shen, L. Xing, S. Ermon. Solving inverse problems in medical imaging with score-based generative models. arXiv preprint arXiv:2111.08005 (2021)
4. Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, B. Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020).
Acknowledgments:
The authors thank Petrobras and Shell, which provided financial support through the Research, Development, and Innovation Investment Clause in collaboration with the Brazilian National Agency of Petroleum, Natural Gas, and Biofuels (ANP, Brazil). Additionally, this research was partially funded by CNPq, CAPES, and FAPERJ.
