Generative Artificial Intelligence for Designing Multi-Scale Hydrogen Fuel Cell Catalyst Layer Nanostructures

Multiscale design of catalyst layers (CLs) is important to advancing hydrogen electrochemical conversion devices toward commercialized deployment, which has nevertheless been greatly hampered by the complex interplay among multiscale CL components, high synthesis cost and vast design space. We lack rational design and optimization techniques that can accurately reflect the nanostructure-performance relationship and cost-effectively search the design space. Here, we fill this gap with a deep generative artificial intelligence (AI) framework, GLIDER, that integrates recent generative AI, data-driven surrogate techniques and collective intelligence to efficiently search the optimal CL nanostructures driven by their electrochemical performance. GLIDER achieves realistic multiscale CL digital generation by leveraging the dimensionality-reduction ability of quantized vector-variational autoencoder. The powerful generative capability of GLIDER allows the efficient search of the optimal design parameters for the Pt-carbon-ionomer nanostructures of CLs. We also demonstrate that GLIDER is transferable to other fuel cell electrode microstructure generation, e.g., fibrous gas diffusion layers and solid oxide fuel cell anode. GLIDER is of potential as a digital tool for the design and optimization of broad electrochemical energy devices.

Figure S6 The accuracy of the surrogate CNN model for predicting J on the training dataset that is applied in Figure 6b.
Supporting Figure S7 3D VQ-VAE model architecture.The modules in front of the Quantized layer composes the 3D encoder.The layers behind the Quantized layer make up the 3D decoder.

Supporting Note 1: The logic CL generation algorithm
The morphology of a typical CL of PEMFC is complicated due to the interaction among metallic Pt nanoparticles (2-3 nm) and amorphous carbon supports, covered by an ionomer film.To generate the digital nano structures of the CL, a stochastic algorithm is developed based on the following assumptions 6 : 1) Carbon supports are assumed as spheres with various diameters (30-70 nm).
2) Secondary pores in carbon supports are excluded.
3) Pt particles are assumed as single cubic boxes with a length of 3.5 nm.4) Pt particles are assumed to attached to the carbon surface.
In the first step, a number of non-overlapped carbon spheres are added to the domain.Then, overlapped carbon spheres are further added until the volume fraction of carbon reaches the target.Subsequentially, Pt particles are attached to the carbon surface randomly until meeting the target Pt loading.The last step is to add ionomer to the above Pt/C nanostructures.The neighbor cells of Pt, carbon and ionomer are randomly assigned as ionomer until the calculated I/C reaches the target.The material properties employed in the algorithm are listed in Table 1.The Python code implemented for the algorithm is available in the section 'Data and code availability'.
During the logic CL generation with varying CL thickness, the volume fraction of each component in CLs is calculated as below: (2) The CL thickness (m) is calculated as follows: where γ Pt (mg cm -2 ) denotes the Pt loading, set as 0.1 and 0.2, respectively, for model validation.
Regarding the logic CL generation with fixed thickness, the CL thickness L and Pt loading γ Pt were set as 1.792 μm and 0.05 mg cm -2 , respectively.The volume fraction of Pt was calculated firstly based on the total Pt mass and Pt density.Then, the volume fraction of carbon, ionomer and pore were resolved according to Eqn. ( 1)-(3).To prepare the training CLs for optimization of the Pt/C and I/C ratios, 600 CLs were generated by randomly assigning Pt/C and I/C ratios from the range 0.2-0.6 and 0.5-1.6,respectively.
(a) The number of used vector labels in 100 test CL samples under three compression ratios f =8, 16 and 32.(b) Slice view of the input and reconstructed CL nanostructures in multiple direction within three cases f =8, 16 and 32.Here, pore (green) regions are displayed for comprehensive comparison.(c) The 3D distributions of the CL reconstruction error for under three compression ratios.(d) Two-point correlation coefficient curves of the real CLs (statistics on 100 samples), and the generated CLs in the case f = 32 (statistics on 100 samples) and f= 8 (statistics on 3 samples due to the prohibitive generation cost).(Supporting Information Page 3) Supporting Figure S2 Evaluation of reconstruction loss when GLIDER compresses GDLs with various compression ratios f =8, 16 and 32.(a) 3D morphology of a real GDL with the size 224×384×384 voxels (resolution 1.33 μm).(b) Real and reconstructed GDLs under various f values.(c) Quantitative analysis of the reconstructed fiber losses against real GDLs.Supporting Figure S3 Comparison between the real (a) and the reconstructed SOFC anode microstructures with a size 128×256×256 voxels (resolution 65 nm) under three sets of compression condition: (b) f d = 4, f h =8 and f v =8; (c) f d = 8, f h =16 and f v =16; (d) f d = 16, f h =32 and f v =32.Supporting Figure S4 Reconstruction of CLs generated by the logic algorithm.The CL is resolved by 512×128×128 voxels (resolution 3.5 nm) under the compression ratio f d = 4, f h =16 and f v =16.(a) training and (b) reconstructed CL.Supporting Figure S5 Generating CLs by transforming 32×32×32 latent under high-efficiency (a) and high-fidelity mode (b).(c) Two-point correlation coefficient curves averaged over 100 generated CLs based on two latent sizes under high-efficiency mode.(d) The generated distributions of Pt/C and I/C ratios of 100 CLs under high-fidelity mode.(e) Four representative CLs in the training dataset.

Table S1
Parameters of the pore-scale CL model.