AI Boosts NOx Reduction Research

A research team from Tianjin University has made significant progress in intelligent catalysis for nitrogen oxide (NOx) reduction, with their latest study published in the prestigious international journal Engineering Applications of Artificial Intelligence. The article, titled "Adaptive extreme learning framework for nitrogen oxide conversion prediction: a multi-scale approach," presents a pioneering artificial intelligence-driven framework for predicting catalyst performance in NOx conversion.

The study was led by Liu Qingling from the School of Environmental Science and Engineering and Yao Mingfa from the State Key Laboratory of Advanced Internal Combustion Engine, with Tong Chengzheng as the first author and Liu Qingling, Liu Caixia, and Yao Mingfa as corresponding authors. The School of Environmental Science and Engineering, Tianjin University, is listed as the first affiliation.

In recent years, artificial intelligence has become a critical driving force for scientific research and interdisciplinary innovation. Tianjin University places high priority on the integration of AI with its strengths in environmental science, materials, and other frontier disciplines. The university has significantly increased its investment in interdisciplinary platforms, advancing the application of AI in cutting-edge scientific research. This study is a representative achievement of the university's “AI+” strategy, underscoring its leadership in intelligent catalysis.

Reducing NOx emissions remains a key challenge in the development of cleaner diesel engines. The design of efficient selective catalytic reduction (SCR) systems has long been hindered by limited experimental data and the complexity of multi-parameter coupling. Traditional machine learning methods struggle to achieve generalizability and computational efficiency in real-world applications.

To address these challenges, the research proposes an Adaptive Extreme Learning Machine (AELM) multi-scale framework, which for the first time integrates automated literature mining with multi-output regression, feature fusion, and temperature-adaptive bias mechanisms. This novel approach enables high-accuracy performance prediction of catalysts under small-sample and highly nonlinear conditions.

The model explicitly characterizes the temperature-conversion relationship using a multi-output regression structure, enhanced by feature weighting and dynamic sampling mechanisms. It significantly improves the model’s capacity to capture complex parameter interactions and temperature dependencies. In tests involving limited data samples, the framework achieved a prediction accuracy improvement of up to 93.3% in NOx conversion. Moreover, the prediction speed increased by approximately 87% compared to conventional neural networks and random forest models, requiring only 0.2–0.3 seconds per inference.

By: Qin Mian