Intelligent Computational Micro-architectured Design of Aerogels for Battery Development

Abstract

Aerogels belong to a class of ultra-light materials with open porous cellular microstructure and extremely low thermal conductivity, which are synthesized by replacing the liquid component of the gel with a gas phase. Due to an interesting combination of these properties, aerogels have often found their application in high temperature insulation and aerospace applications, most famously silica aerogels in the stardust collector by NASA. Recently, sulfur infiltrated carbon aerogels as the cathode, have been proven to be an effective alternate to the existing lithium sulfur batteries, with increased charge capacities [1].

The carbon aerogels are synthesized by the pyrolysis of organic aerogels e.g. resorcinol-formaldehyde (RF) aerogels. However, it’s development with desired permeability for effective sulfur infiltration is a challenge due to the iterative nature of the synthesis process. Thus, the development of digital twins could help accelerate the optimization and development time for aerogels with desired material properties. In this contribution, an artificial intelligence-assisted modeling technique is presented for architectured microstructure design of RF aerogels with desired permeability. The morphology of RF aerogel’s representative volume element (RVE) was shown to be modeled with a 3-parameter Gaussian random field algorithm (GRF) [2]. By applying a similar model strategy the absolute permeability of the computational models is estimated by means of a watershed image segmentation algorithm to develop pore network models (PNM). The PNMs help analyse the pore size distribution and the infiltration characteristics through the RVE (the maximum flow path and the absolute permeability). For the rapid analysis and reverse engineering of aerogels with desired microstructures, a machine learning framework is developed. The environment constitutes of a surrogate model (artificial neural network) that maps the GRF parameters to the aerogel morphology and the infiltration characteristics. The surrogate model thus acts as an intelligent digital twin, eliminating the need for iterative computational modeling and its post-processing for permeability analysis. Furthermore, in the environment, a reinforcement learning agent (deep deterministic policy gradient agent) works in combination with the surrogate model to optimise the parameters of the GRF to achieve the target material infiltration characteristics. This enables the prediction of the RF aerogel’s material properties without extensive and iterative computational modeling, thus accelerating the process of aerogel development.

References

1. M. Nojabaee et al., J. Mater. Chem. A 9, 6508-6519 (2021)

2. C.J. Gommes et al., Phys. Rev. E 77, 041409 (2018)