Modular devices such as electrolyzers, fuel cells, and flow batteries are essential for renewable energy conversion. However, scaling up from individual reaction cells to commercial-scale systems poses significant challenges in maintaining operational consistency due to non-uniform distributions of flow, mass, and heat among numerous parallel units. In a study published in ENG. Chem. Eng., researchers at East China University of Science and Technology and Suzhou National Laboratory propose a data-driven framework to uncover design rules for modular CO₂ reduction reactors.
The team generated two data sets: a random architecture data set of approximately 5,000 samples using Latin hypercube sampling, and a Pareto-optimal data set via multi-objective optimization maximizing CO₂ conversion and net CO₂ consumption while minimizing total annual cost. XGBoost regression models were trained to map structural features to performance metrics, achieving R² values of 0.90–0.99 on test sets, enabling accurate prediction and interpretation.
SHAP analysis revealed that different performance objectives are governed by distinct architectural hierarchies. Total annual cost is primarily determined by substrate-level features: the number of reaction channels per substrate and the size of on-substrate distribution channels. A high channel density with an optimized distributor size minimizes material and operating costs. In contrast, CO₂ conversion and net carbon footprint are dictated at the stack level. The number of stacks and inter-stack connection schemes are most influential. For optimal reaction performance, fewer than 10 parallel stacks with sufficiently large inter-stack distribution pipelines are recommended. Series connections should be avoided as they intensify intra-stack maldistribution and energy consumption.
An XGBoost classification model was trained on the combined data set to distinguish Pareto-optimal designs from non-optimal ones, achieving an F₁ score of 0.97. A simple heuristic emerged: architectures with 100 to 200 substrates per stack, on-substrate distribution channels smaller than 5.4 mm, and at least 25 reaction channels per substrate have over 90 % likelihood of being Pareto-optimal.
Unsupervised learning via PCA and K-means clustering identified four basic Pareto-optimal patterns: small-sized single stacks, few small-sized stacks in parallel, large-sized single stacks, and few large-sized stacks in parallel. The large-sized few-stack parallel pattern achieves the best reaction conversion (67.1 %) due to larger distributors and balanced branching, while the small-sized single-stack pattern offers the lowest total annual cost (14,832 CNY·a⁻¹) at the expense of conversion (60.1 %).
Scenario-specific analysis revealed that increasing production scale necessitates larger-sized modules to maintain distribution uniformity, while high feedstock volatility favors highly parallelized architectures with many stacks to buffer inlet flow fluctuations.
This work provides a generalizable paradigm for understanding design principles and accelerating the development of next-generation modular chemical manufacturing systems.

DOI
10.1007/s11705-026-2678-y