Deep Learning Kinetic Modeling of the Catalytic Decomposition of Ammonia in Green Hydrogen Production: Effects of Catalyst Composition and Operating Variables

The increasing demand
for sustainable energy has accelerated the
development of green hydrogen production technologies. Among these,
the catalytic decomposition of ammonia stands out because of its efficient
storage and transportation as well as its compatibility with existing
infrastructure. Nevertheless, challenges in enhancing the reaction
performance still hinder its large-scale implementation. To address
these limitations and optimize the process, this work presents the
development of a deep-learning-based artificial neural network to
model ammonia conversion as a function of operating conditions and
catalyst composition encoded directly into the network. The final
model designed significantly outperformed traditional machine learning
techniques and the smaller architectures tested. Deep learning was
fundamental for achieving the lowest predictive errors (RMSE = 10.06
ppm and MAE = 7.98 ppm) and minimizing the difference between training
and validation errors, indicating a high degree of stability and generalization.
A comprehensive sensitivity analysis was also conducted and aligned
with literature findings, revealing the model’s capacity to
capture complex physicochemical patterns. Finally, validation on external
data further confirmed its generalization capabilities. To the best
of our knowledge, this is the first study to implement deep neural
networks for modeling the catalytic decomposition of ammonia, including
catalyst compositional features, while also contributing to the broader,
still-emerging application of deep learning in catalytic systems.

全球可持续能源需求的持续攀升，推动了绿氢生产技术的快速发展。其中，氨催化分解技术凭借高效的储运特性以及与现有工业基础设施的良好兼容性，成为该领域的佼佼者。然而，提升反应性能所面临的诸多挑战，仍制约着该技术的大规模工业化落地。为解决上述局限并优化工艺流程，本研究开发了一款基于深度学习的人工神经网络（Artificial Neural Network, ANN），将操作条件与催化剂组成作为输入变量直接编码至网络结构中，以实现对氨转化率的精准建模。所构建的最终模型性能显著优于传统机器学习方法与测试过的小型网络架构。深度学习是实现最低预测误差（均方根误差（Root Mean Square Error, RMSE）=10.06 ppm，平均绝对误差（Mean Absolute Error, MAE）=7.98 ppm）、最小化训练与验证误差差值的核心支撑，这表明模型具备极高的稳定性与泛化能力。本研究还开展了全面的敏感性分析，结果与已有文献结论高度吻合，证明该模型能够有效捕捉复杂的物理化学过程规律。最后，通过外部独立数据集的验证，进一步证实了模型的泛化性能。据我们所知，本研究是首个将深度神经网络应用于包含催化剂组成特征的氨催化分解建模的研究，同时也为深度学习在催化系统这一仍处于发展初期的新兴应用领域提供了有益的实践参考。